Upload 15 files

LICENSE

@@ -0,0 +1,395 @@
+Attribution 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public:
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution 4.0 International Public License ("Public License"). To the
+extent this Public License may be interpreted as a contract, You are
+granted the Licensed Rights in consideration of Your acceptance of
+these terms and conditions, and the Licensor grants You such rights in
+consideration of benefits the Licensor receives from making the
+Licensed Material available under these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  j. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  k. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part; and
+
+            b. produce, reproduce, and Share Adapted Material.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

environment.yml

@@ -0,0 +1,271 @@
+name: mediq
+channels:
+  - pytorch
+  - nvidia
+  - conda-forge
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=conda_forge
+  - _openmp_mutex=4.5=2_gnu
+  - aiohappyeyeballs=2.4.4=py312h06a4308_0
+  - aiohttp=3.10.5=py312h5eee18b_0
+  - aiosignal=1.2.0=pyhd3eb1b0_0
+  - annotated-types=0.6.0=py312h06a4308_0
+  - anyio=4.6.2=py312h06a4308_0
+  - attrs=24.2.0=py312h06a4308_0
+  - aws-c-auth=0.7.22=h96bc93b_2
+  - aws-c-cal=0.6.14=h88a6e22_1
+  - aws-c-common=0.9.19=h4ab18f5_0
+  - aws-c-compression=0.2.18=h83b837d_6
+  - aws-c-event-stream=0.4.2=ha47c788_12
+  - aws-c-http=0.8.1=h29d6fba_17
+  - aws-c-io=0.14.8=h21d4f22_5
+  - aws-c-mqtt=0.10.4=h759edc4_4
+  - aws-c-s3=0.5.9=h594631b_3
+  - aws-c-sdkutils=0.1.16=h83b837d_2
+  - aws-checksums=0.1.18=h83b837d_6
+  - aws-crt-cpp=0.26.9=he3a8b3b_0
+  - aws-sdk-cpp=1.11.329=hba8bd5f_3
+  - blas=1.0=mkl
+  - bottleneck=1.4.2=py312ha883a20_0
+  - brotli-python=1.0.9=py312h6a678d5_8
+  - bzip2=1.0.8=h5eee18b_6
+  - c-ares=1.34.4=hb9d3cd8_0
+  - ca-certificates=2024.12.31=h06a4308_0
+  - certifi=2024.12.14=py312h06a4308_0
+  - charset-normalizer=3.3.2=pyhd3eb1b0_0
+  - cuda-cudart=12.4.127=0
+  - cuda-cupti=12.4.127=0
+  - cuda-libraries=12.4.1=0
+  - cuda-nvrtc=12.4.127=0
+  - cuda-nvtx=12.4.127=0
+  - cuda-opencl=12.6.77=0
+  - cuda-runtime=12.4.1=0
+  - cuda-version=12.6=3
+  - datasets=3.2.0=pyhd8ed1ab_0
+  - dill=0.3.8=py312h06a4308_0
+  - distro=1.9.0=py312h06a4308_0
+  - expat=2.6.4=h6a678d5_0
+  - ffmpeg=4.3=hf484d3e_0
+  - freetype=2.12.1=h4a9f257_0
+  - frozenlist=1.5.0=py312h5eee18b_0
+  - fsspec=2024.6.1=py312h06a4308_0
+  - gflags=2.2.2=h6a678d5_1
+  - giflib=5.2.2=h5eee18b_0
+  - glog=0.7.1=hbabe93e_0
+  - gmp=6.2.1=h295c915_3
+  - gnutls=3.6.15=he1e5248_0
+  - h11=0.14.0=py312h06a4308_0
+  - httpcore=1.0.2=py312h06a4308_0
+  - httpx=0.27.0=py312h06a4308_0
+  - huggingface_hub=0.24.6=py312h06a4308_0
+  - idna=3.7=py312h06a4308_0
+  - intel-openmp=2023.1.0=hdb19cb5_46306
+  - jinja2=3.1.4=py312h06a4308_1
+  - jiter=0.6.1=py312hb02cf49_0
+  - jpeg=9e=h5eee18b_3
+  - krb5=1.20.1=h143b758_1
+  - lame=3.100=h7b6447c_0
+  - lcms2=2.16=hb9589c4_0
+  - ld_impl_linux-64=2.40=h12ee557_0
+  - lerc=4.0.0=h6a678d5_0
+  - libabseil=20240116.2=cxx17_h6a678d5_0
+  - libarrow=16.1.0=hcb6531f_6_cpu
+  - libarrow-acero=16.1.0=hac33072_6_cpu
+  - libarrow-dataset=16.1.0=hac33072_6_cpu
+  - libarrow-substrait=16.1.0=h7e0c224_6_cpu
+  - libbrotlicommon=1.1.0=hb9d3cd8_2
+  - libbrotlidec=1.1.0=hb9d3cd8_2
+  - libbrotlienc=1.1.0=hb9d3cd8_2
+  - libcrc32c=1.1.2=h6a678d5_0
+  - libcublas=12.4.5.8=0
+  - libcufft=11.2.1.3=0
+  - libcufile=1.11.1.6=0
+  - libcurand=10.3.7.77=0
+  - libcurl=8.9.1=h251f7ec_0
+  - libcusolver=11.6.1.9=0
+  - libcusparse=12.3.1.170=0
+  - libdeflate=1.22=h5eee18b_0
+  - libedit=3.1.20230828=h5eee18b_0
+  - libev=4.33=h7f8727e_1
+  - libevent=2.1.12=hdbd6064_1
+  - libexpat=2.6.4=h5888daf_0
+  - libffi=3.4.4=h6a678d5_1
+  - libgcc=14.2.0=h77fa898_1
+  - libgcc-ng=14.2.0=h69a702a_1
+  - libgomp=14.2.0=h77fa898_1
+  - libgoogle-cloud=2.24.0=h2736e30_0
+  - libgoogle-cloud-storage=2.24.0=h3d9a0c8_0
+  - libgrpc=1.62.2=h15f2491_0
+  - libiconv=1.16=h5eee18b_3
+  - libidn2=2.3.4=h5eee18b_0
+  - libjpeg-turbo=2.0.0=h9bf148f_0
+  - libnghttp2=1.57.0=h2d74bed_0
+  - libnpp=12.2.5.30=0
+  - libnsl=2.0.1=hd590300_0
+  - libnvfatbin=12.6.77=0
+  - libnvjitlink=12.4.127=0
+  - libnvjpeg=12.3.1.117=0
+  - libparquet=16.1.0=h6a7eafb_6_cpu
+  - libpng=1.6.39=h5eee18b_0
+  - libprotobuf=4.25.3=he621ea3_0
+  - libre2-11=2023.09.01=h5a48ba9_2
+  - libsqlite=3.46.0=hde9e2c9_0
+  - libssh2=1.11.1=h251f7ec_0
+  - libstdcxx=14.2.0=hc0a3c3a_1
+  - libstdcxx-ng=14.2.0=h4852527_1
+  - libtasn1=4.19.0=h5eee18b_0
+  - libthrift=0.19.0=hb90f79a_1
+  - libtiff=4.5.1=hffd6297_1
+  - libunistring=0.9.10=h27cfd23_0
+  - libutf8proc=2.8.0=hf23e847_1
+  - libuuid=2.38.1=h0b41bf4_0
+  - libwebp=1.3.2=h11a3e52_0
+  - libwebp-base=1.3.2=h5eee18b_1
+  - libxcrypt=4.4.36=hd590300_1
+  - libzlib=1.2.13=h4ab18f5_6
+  - llvm-openmp=14.0.6=h9e868ea_0
+  - lz4-c=1.9.4=h6a678d5_1
+  - markupsafe=2.1.3=py312h5eee18b_0
+  - mkl=2023.1.0=h213fc3f_46344
+  - mkl-service=2.4.0=py312h5eee18b_1
+  - mkl_fft=1.3.11=py312h5eee18b_0
+  - mkl_random=1.2.8=py312h526ad5a_0
+  - mpmath=1.3.0=py312h06a4308_0
+  - multidict=6.1.0=py312h5eee18b_0
+  - multiprocess=0.70.15=py312h06a4308_0
+  - ncurses=6.4=h6a678d5_0
+  - nettle=3.7.3=hbbd107a_1
+  - networkx=3.2.1=py312h06a4308_0
+  - numexpr=2.10.1=py312h3c60e43_0
+  - openai=1.57.4=pyhd8ed1ab_1
+  - openh264=2.1.1=h4ff587b_0
+  - openjpeg=2.5.2=he7f1fd0_0
+  - openssl=3.4.0=hb9d3cd8_0
+  - orc=2.0.1=h2d29ad5_0
+  - packaging=24.2=py312h06a4308_0
+  - pandas=2.2.3=py312h6a678d5_0
+  - pip=24.2=py312h06a4308_0
+  - propcache=0.2.0=py312h5eee18b_0
+  - pyarrow=16.1.0=py312h9cebb41_2
+  - pyarrow-core=16.1.0=py312h0983c49_2_cpu
+  - pysocks=1.7.1=py312h06a4308_0
+  - python=3.12.2=hab00c5b_0_cpython
+  - python-dateutil=2.9.0post0=py312h06a4308_2
+  - python-tzdata=2023.3=pyhd3eb1b0_0
+  - python-xxhash=2.0.2=py312h5eee18b_1
+  - python_abi=3.12=5_cp312
+  - pytorch=2.5.1=py3.12_cuda12.4_cudnn9.1.0_0
+  - pytorch-cuda=12.4=hc786d27_7
+  - pytorch-mutex=1.0=cuda
+  - pytz=2024.1=py312h06a4308_0
+  - pyyaml=6.0.2=py312h5eee18b_0
+  - re2=2023.09.01=h7f4b329_2
+  - readline=8.2=h5eee18b_0
+  - regex=2024.9.11=py312h5eee18b_0
+  - requests=2.32.3=py312h06a4308_1
+  - s2n=1.4.15=he19d79f_0
+  - safetensors=0.4.5=py312hc50d6dc_1
+  - setuptools=75.1.0=py312h06a4308_0
+  - six=1.16.0=pyhd3eb1b0_1
+  - snappy=1.2.1=h6a678d5_0
+  - sniffio=1.3.0=py312h06a4308_0
+  - sqlite=3.45.3=h5eee18b_0
+  - tbb=2021.8.0=hdb19cb5_0
+  - tk=8.6.14=h39e8969_0
+  - tokenizers=0.21.0=py312h8360d73_0
+  - torchaudio=2.5.1=py312_cu124
+  - torchtriton=3.1.0=py312
+  - torchvision=0.20.1=py312_cu124
+  - tqdm=4.66.5=py312he106c6f_0
+  - transformers=4.47.1=pyhd8ed1ab_0
+  - tzdata=2024b=h04d1e81_0
+  - urllib3=2.2.3=py312h06a4308_0
+  - vllm-nccl-cu12=2.18.1.0.4.0=pyh52da0d0_1
+  - wheel=0.44.0=py312h06a4308_0
+  - xxhash=0.8.0=h7f8727e_3
+  - xz=5.4.6=h5eee18b_1
+  - yaml=0.2.5=h7b6447c_0
+  - yarl=1.18.0=py312h5eee18b_0
+  - zlib=1.2.13=h4ab18f5_6
+  - zstd=1.5.6=hc292b87_0
+  - pip:
+      - aiohttp-cors==0.7.0
+      - airportsdata==20241001
+      - astor==0.8.1
+      - blake3==1.0.4
+      - cachetools==5.5.1
+      - click==8.1.8
+      - cloudpickle==3.1.1
+      - colorful==0.5.6
+      - compressed-tensors==0.8.1
+      - depyf==0.18.0
+      - diskcache==5.6.3
+      - distlib==0.3.9
+      - einops==0.8.0
+      - fastapi==0.115.7
+      - filelock==3.17.0
+      - gguf==0.10.0
+      - google-api-core==2.24.0
+      - google-auth==2.38.0
+      - googleapis-common-protos==1.66.0
+      - grpcio==1.70.0
+      - httptools==0.6.4
+      - importlib-metadata==8.6.1
+      - iniconfig==2.0.0
+      - interegular==0.3.3
+      - jsonschema==4.23.0
+      - jsonschema-specifications==2024.10.1
+      - lark==1.2.2
+      - lm-format-enforcer==0.10.9
+      - mistral-common==1.5.2
+      - msgpack==1.1.0
+      - msgspec==0.19.0
+      - nest-asyncio==1.6.0
+      - numpy==1.26.4
+      - nvidia-ml-py==12.570.86
+      - opencensus==0.11.4
+      - opencensus-context==0.1.3
+      - opencv-python-headless==4.11.0.86
+      - outlines==0.1.11
+      - outlines-core==0.1.26
+      - partial-json-parser==0.2.1.1.post5
+      - pillow==10.4.0
+      - platformdirs==4.3.6
+      - pluggy==1.5.0
+      - prometheus-client==0.21.1
+      - prometheus-fastapi-instrumentator==7.0.2
+      - proto-plus==1.25.0
+      - protobuf==5.29.3
+      - psutil==6.1.1
+      - py-cpuinfo==9.0.0
+      - py-spy==0.4.0
+      - pyasn1==0.6.1
+      - pyasn1-modules==0.4.1
+      - pybind11==2.13.6
+      - pycountry==24.6.1
+      - pydantic==2.10.6
+      - pydantic-core==2.27.2
+      - pytest==8.3.4
+      - python-dotenv==1.0.1
+      - pyzmq==26.2.0
+      - ray==2.41.0
+      - referencing==0.36.2
+      - rpds-py==0.22.3
+      - rsa==4.9
+      - sentencepiece==0.2.0
+      - smart-open==7.1.0
+      - starlette==0.45.3
+      - sympy==1.13.1
+      - tiktoken==0.7.0
+      - typing-extensions==4.12.2
+      - uvicorn==0.34.0
+      - uvloop==0.21.0
+      - virtualenv==20.29.1
+      - vllm==0.6.6.post1
+      - watchfiles==1.0.4
+      - websockets==14.2
+      - wrapt==1.17.2
+      - xformers==0.0.28.post3
+      - xgrammar==0.1.11
+      - zipp==3.21.0

readme.md

@@ -0,0 +1,79 @@
+# MediQ: Question-Asking LLMs for Adaptive and Reliable Clinical Reasoning
+
+## [[paper](https://arxiv.org/abs/2406.00922)] [[website](https://stellalisy.com/projects/mediQ/)] [[data](https://github.com/stellali7/mediQ/tree/main/data)]
+
+## Overview
+This benchmark system simulates an interactive conversation between a patient and an expert. The system evaluates how well participants' expert modules can handle realistic patient queries by either asking relevant questions or making final decisions based on the conversation history.
+
+## Installation
+Clone this repository to your local machine using the following command:
+```
+git clone https://github.com/stellali7/MediQ.git
+```
+
+Navigate into the project directory:
+```
+cd MediQ
+```
+
+Create a new conda environment with necessary packages (note: you need to be on a GPU node to install PyTorch with CUDA):
+```
+conda env create -f environment.yml
+```
+
+
+## Project Structure
+- `benchmark.py`: Main script to run the benchmark.
+- `patient.py`: Defines the `Patient` class that simulates patient behavior.
+- `expert.py`: Contains the `Expert` class which participants will extend to implement their response strategies.
+- `args.py`: Handles command-line arguments for the benchmark system.
+
+## Configuration
+Before running the benchmark, configure the necessary parameters in `args.py`:
+- `--expert_module`: The file name (without `.py`) where the Expert class is implemented (e.g. expert if your Expert class definition is in `expert.py`)
+- `--expert_class`: The name of the Expert class to be evaluated, this should be defined in the file `[expert_module].py` (e.g. RandomExpert)
+- `--patient_module`: The file name (without `.py`) where the Patient class is implemented (e.g. patient if your Patient class definition is in `patient.py`)
+- `--patient_class`: The name of the Patient class to use for the benchmark, this should be defined in the file `[patient_module].py` (e.g. RandomPatient)
+- `--data_dir`: Directory containing the development data files.
+- `--dev_filename`: Filename for development data.
+- `--log_filename`: Filename for logging general benchmark information.
+- `--history_log_filename`: Filename for logging detailed interaction history.
+- `--message_log_filename`: Filename for logging messages.
+- `--output_filepath`: Path where the output JSONL files will be saved.
+
+## Running the Benchmark
+NOTE: if you choose to use an OpenAI model to power the benchmark, you need to put the API key in `src/keys.py`.
+
+To test run the benchmark, use the following command (note: the Patient system is provided as described in the paper, the Expert system is a skeleton code. For a fast test run, use `--patient_variant random` to not call use any actual model or API):
+```
+python mediQ_benchmark.py  --expert_module expert --expert_class FixedExpert \
+                        --patient_module patient --patient_class RandomPatient \
+                        --data_dir ../data --dev_filename all_dev_good.jsonl \
+                        --output_filename out.jsonl --max_questions 10
+```
+
+Ensure to replace the placeholder values with actual parameters relevant to your setup.
+
+## Try out your own Expert system
+You can easily create their own `Expert` class within a module specified by `--expert_module`, or old a different model by specifying the model path in `--expert_model`. The class should correctly implement the `respond` method to interact with the `Patient` instances based on their states (the Patient can be customized as well). The response should either be a continuation question or a final decision. Your implementation will be tested against a variety of patient scenarios provided in the development dataset.
+
+## How to Cite
+```
+@inproceedings{li2024mediq,
+        title={MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning},
+        author={Li, Shuyue Stella and Balachandran, Vidhisha and Feng, Shangbin and Ilgen, Jonathan S and Pierson, Emma and Koh, Pang Wei and Tsvetkov, Yulia},
+        journal={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
+        year={2024}
+      }
+```
+
+Shield: [![CC BY 4.0][cc-by-shield]][cc-by]
+
+This work is licensed under a
+[Creative Commons Attribution 4.0 International License][cc-by].
+
+[![CC BY 4.0][cc-by-image]][cc-by]
+
+[cc-by]: http://creativecommons.org/licenses/by/4.0/
+[cc-by-image]: https://i.creativecommons.org/l/by/4.0/88x31.png
+[cc-by-shield]: https://img.shields.io/badge/License-CC%20BY%204.0-lightgrey.svg

src/args.py

@@ -0,0 +1,46 @@
+import argparse
+import os
+
+def get_args():
+    parser = argparse.ArgumentParser(description="Run the benchmark with specified configurations.")
+    parser.add_argument('--expert_module', type=str, default='expert', help='file name where the expert class is implemented.')
+    parser.add_argument('--expert_class', type=str, required=True, help='Expert class name to use for the benchmark.')
+    parser.add_argument('--expert_model', type=str, default='meta-llama/Llama-3.1-8B-Instruct', help='Expert model name to use for the benchmark, can be a local model or a Huggingface model.')
+    parser.add_argument('--expert_model_question_generator', type=str, default='meta-llama/Llama-3.1-8B-Instruct', help='You can set a separate model for the follow-up question generator, can be a local model or a Huggingface model.')
+    
+    parser.add_argument('--patient_module', type=str, default='patient', help='file name where the patient class is implemented.')
+    parser.add_argument('--patient_class', type=str, required=True, help='Patient class name to use for the benchmark.')
+    parser.add_argument('--patient_model', type=str, default='meta-llama/Llama-3.1-8B-Instruct', help='Patient model name to use for the benchmark, can be a local model or a Huggingface model.')
+    
+    parser.add_argument('--data_dir', type=str, required=True, help='Directory containing the development data files.')
+    parser.add_argument('--dev_filename', type=str, required=True, help='Filename for development data.')
+
+    parser.add_argument('--output_filename', type=str, default="results.jsonl")
+
+    parser.add_argument("--max_questions", type=int, default=30)
+
+    parser.add_argument('--log_filename', type=str, default='log.log', help='Filename for logging general benchmark results.')
+    parser.add_argument('--history_log_filename', type=str, default=None, help='Filename for logging interaction history, will not log if None.')
+    parser.add_argument('--detail_log_filename', type=str, default=None, help='Filename for logging detailed prompts and response on abstention, will not log if None.')
+    parser.add_argument('--message_log_filename', type=str, default=None, help='Filename for logging messages passed into API calls, will not log if None.')
+
+    parser.add_argument('--rationale_generation', action='store_true', help='Generate rationales for the choices.')
+    parser.add_argument('--self_consistency', type=int, default=1, help='Number of times to run the self-consistency check.')
+    parser.add_argument('--abstain_threshold', type=float, default=0.8, help='Threshold for abstaining from making a choice.')
+    parser.add_argument('--independent_modules', action='store_true', help='Cognitive modules within the Expert dont see previous convo.')
+
+    parser.add_argument('--use_vllm', action='store_true', help='Use the VLLM model for generating responses.')
+    parser.add_argument('--use_api', type=str, default=None, help='Use an API for generating responses.', choices=['openai']) # compatible with the OpenAI API for now
+    parser.add_argument('--temperature', type=float, default=0.6, help='Temperature for sampling from the model.')
+    parser.add_argument('--top_p', type=float, default=0.9, help='Top p value for nucleus sampling.')
+    parser.add_argument('--max_tokens', type=int, default=256, help='Maximum number of tokens to generate.')
+    parser.add_argument('--top_logprobs', type=int, default=0, help='Number of top logprobs to return.')
+    parser.add_argument('--api_account', type=str, default="mediQ", help='API keys are stored in keys.py, api_account is the name of the key.')
+    
+    args =  parser.parse_args()
+
+    if args.log_filename: os.makedirs(os.path.dirname(args.log_filename), exist_ok=True)
+    if args.history_log_filename: os.makedirs(os.path.dirname(args.history_log_filename), exist_ok=True)
+    if args.detail_log_filename: os.makedirs(os.path.dirname(args.detail_log_filename), exist_ok=True)
+    if args.message_log_filename: os.makedirs(os.path.dirname(args.message_log_filename), exist_ok=True)
+    return args

src/evaluate.py

@@ -0,0 +1,56 @@
+import json
+import numpy as np
+import torch
+from sentence_transformers import SentenceTransformer, util
+
+# from mydifflib import get_close_matches
+
+device = torch.device("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() and torch.backends.mps.is_built() else "cpu")
+print(f"Device: {device}")
+
+emb_model = SentenceTransformer('stsb-roberta-large', device=device)
+
+def eval_sample(id, sample, choice, scores, questions, answers, answer_dne, temp_choice_list, threshold=0.85):
+    questions_emb = emb_model.encode(questions)
+    facts_emb = emb_model.encode(sample["facts"])
+    facts_count = [0]*len(sample["facts"])
+    answers_expanded, answers_count = [], []
+    
+    for answer in answers:
+        answer = [a for a in answer.split('. ') if not a.isnumeric()]  # split the answer into atomic facts
+        answers_expanded.extend(answer)
+        answers_count.append(len(answer))
+    answers_emb = emb_model.encode(answers_expanded)
+
+    output_dict = {
+        "id": id,
+        "info": sample,
+        "interactive_system": {
+            "choice": choice,
+            "confidence_scores": scores,
+            "questions": questions,
+            "answers": answers,
+            "answer_dne": answer_dne,
+            "num_questions": len(questions),
+            "intermediate_choices": temp_choice_list,
+        },
+        "eval": {
+            "repeat_question_score": [],
+            "repeat_answer_score": [],
+            "relevancy_score": [],
+            "delta_confidence_score": [],
+            "specificity_score": []
+        }
+    }
+    
+    # Example placeholder for evaluation metrics computation
+    for i in range(len(questions)):
+        output_dict["eval"]["repeat_question_score"].append(np.random.random())  # Placeholder
+        output_dict["eval"]["repeat_answer_score"].append(np.random.random())  # Placeholder
+        output_dict["eval"]["relevancy_score"].append(np.random.random())  # Placeholder
+        output_dict["eval"]["delta_confidence_score"].append(np.random.random())  # Placeholder
+        output_dict["eval"]["specificity_score"].append(np.random.random())  # Placeholder
+
+    return output_dict
+
+# Other functions should be similarly reviewed and implemented

src/expert.py

@@ -0,0 +1,225 @@
+import random
+import expert_functions
+
+class Expert:
+    """
+    Expert system skeleton
+    """
+    def __init__(self, args, inquiry, options):
+        # Initialize the expert with necessary parameters and the initial context or inquiry
+        self.args = args
+        self.inquiry = inquiry
+        self.options = options
+
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        raise NotImplementedError
+    
+    def ask_question(self, patient_state, prev_messages):
+        # Generate a question based on the current patient state
+        kwargs = {
+            "patient_state": patient_state,
+            "inquiry": self.inquiry,
+            "options_dict": self.options,
+            "messages": prev_messages,
+            "independent_modules": self.args.independent_modules,
+            "model_name": self.args.expert_model_question_generator,
+            "use_vllm": self.args.use_vllm,
+            "use_api": self.args.use_api,
+            "temperature": self.args.temperature,
+            "max_tokens": self.args.max_tokens,
+            "top_p": self.args.top_p,
+            "top_logprobs": self.args.top_logprobs,
+            "api_account": self.args.api_account
+        }
+        return expert_functions.question_generation(**kwargs)
+    
+    def get_abstain_kwargs(self, patient_state):
+        kwargs = {
+            "max_depth": self.args.max_questions,
+            "patient_state": patient_state,
+            "rationale_generation": self.args.rationale_generation,
+            "inquiry": self.inquiry,
+            "options_dict": self.options,
+            "abstain_threshold": self.args.abstain_threshold,
+            "self_consistency": self.args.self_consistency,
+            "model_name": self.args.expert_model,
+            "use_vllm": self.args.use_vllm,
+            "use_api": self.args.use_api,
+            "temperature": self.args.temperature,
+            "max_tokens": self.args.max_tokens,
+            "top_p": self.args.top_p,
+            "top_logprobs": self.args.top_logprobs,
+            "api_account": self.args.api_account
+        }
+        return kwargs
+
+
+class RandomExpert(Expert):
+    """
+    Below is an example Expert system that randomly asks a question or makes a choice based on the current patient state.
+    This should be replaced with a more sophisticated expert system that can make informed decisions based on the patient state.
+    """
+
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        initial_info = patient_state['initial_info']  # not use because it's random
+        history = patient_state['interaction_history']  # not use because it's random
+
+        # randomly decide to ask a question or make a choice
+        abstain = random.random() < 0.5
+        toy_question = "Can you describe your symptoms more?"
+        toy_decision = self.choice(patient_state)
+        conf_score = random.random()/2 if abstain else random.random()
+
+        return {
+            "type": "question" if abstain else "choice",
+            "question": toy_question,
+            "letter_choice": toy_decision,
+            "confidence": conf_score,  # Optional confidence score
+            "urgent": True,  # Example of another optional flag
+            "additional_info": "Check for any recent changes."  # Any other optional data
+        }
+
+    def choice(self, patient_state):
+        # Generate a choice or intermediate decision based on the current patient state
+        # randomly choose an option
+        return random.choice(list(self.options.keys()))
+
+
+class BasicExpert(Expert):
+    def respond(self, patient_state):
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.implicit_abstention_decision(**kwargs)
+        return {
+            "type": "question" if abstain_response_dict["abstain"] else "choice",
+            "question": abstain_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }
+
+
+class FixedExpert(Expert):
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.fixed_abstention_decision(**kwargs)
+        if abstain_response_dict["abstain"] == False:
+            return {
+                "type": "choice",
+                "letter_choice": abstain_response_dict["letter_choice"],
+                "confidence": abstain_response_dict["confidence"],
+                "usage": abstain_response_dict["usage"]
+            }
+
+        question_response_dict = self.ask_question(patient_state, abstain_response_dict["messages"])
+        abstain_response_dict["usage"]["input_tokens"] += question_response_dict["usage"]["input_tokens"]
+        abstain_response_dict["usage"]["output_tokens"] += question_response_dict["usage"]["output_tokens"]
+        return {
+            "type": "question",
+            "question": question_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }
+        
+
+class BinaryExpert(Expert):
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.binary_abstention_decision(**kwargs)
+        if abstain_response_dict["abstain"] == False:
+            return {
+                "type": "choice",
+                "letter_choice": abstain_response_dict["letter_choice"],
+                "confidence": abstain_response_dict["confidence"],
+                "usage": abstain_response_dict["usage"]
+            }
+
+        question_response_dict = self.ask_question(patient_state, abstain_response_dict["messages"])
+        abstain_response_dict["usage"]["input_tokens"] += question_response_dict["usage"]["input_tokens"]
+        abstain_response_dict["usage"]["output_tokens"] += question_response_dict["usage"]["output_tokens"]
+        return {
+            "type": "question",
+            "question": question_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }
+
+
+class NumericalExpert(Expert):
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.numerical_abstention_decision(**kwargs)
+        if abstain_response_dict["abstain"] == False:
+            return {
+                "type": "choice",
+                "letter_choice": abstain_response_dict["letter_choice"],
+                "confidence": abstain_response_dict["confidence"],
+                "usage": abstain_response_dict["usage"]
+            }
+
+        question_response_dict = self.ask_question(patient_state, abstain_response_dict["messages"])
+        abstain_response_dict["usage"]["input_tokens"] += question_response_dict["usage"]["input_tokens"]
+        abstain_response_dict["usage"]["output_tokens"] += question_response_dict["usage"]["output_tokens"]
+        return {
+            "type": "question",
+            "question": question_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }
+
+
+class NumericalCutOffExpert(Expert):
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.numcutoff_abstention_decision(**kwargs)
+        if abstain_response_dict["abstain"] == False:
+            return {
+                "type": "choice",
+                "letter_choice": abstain_response_dict["letter_choice"],
+                "confidence": abstain_response_dict["confidence"],
+                "usage": abstain_response_dict["usage"]
+            }
+
+        question_response_dict = self.ask_question(patient_state, abstain_response_dict["messages"])
+        abstain_response_dict["usage"]["input_tokens"] += question_response_dict["usage"]["input_tokens"]
+        abstain_response_dict["usage"]["output_tokens"] += question_response_dict["usage"]["output_tokens"]
+        return {
+            "type": "question",
+            "question": question_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }
+
+
+class ScaleExpert(Expert):
+    def respond(self, patient_state):
+        # Decision-making based on the initial information, history of interactions, current inquiry, and options
+        kwargs = self.get_abstain_kwargs(patient_state)
+        abstain_response_dict = expert_functions.scale_abstention_decision(**kwargs)
+        if abstain_response_dict["abstain"] == False:
+            return {
+                "type": "choice",
+                "letter_choice": abstain_response_dict["letter_choice"],
+                "confidence": abstain_response_dict["confidence"],
+                "usage": abstain_response_dict["usage"]
+            }
+
+        question_response_dict = self.ask_question(patient_state, abstain_response_dict["messages"])
+        abstain_response_dict["usage"]["input_tokens"] += question_response_dict["usage"]["input_tokens"]
+        abstain_response_dict["usage"]["output_tokens"] += question_response_dict["usage"]["output_tokens"]
+        return {
+            "type": "question",
+            "question": question_response_dict["atomic_question"],
+            "letter_choice": abstain_response_dict["letter_choice"],
+            "confidence": abstain_response_dict["confidence"],
+            "usage": abstain_response_dict["usage"]
+        }

src/expert_basics.py

@@ -0,0 +1,305 @@
+import logging
+import random
+import re
+from helper import get_response
+
+
+def log_info(message, logger_name="detail_logger", print_to_std=False, type="info"):
+    # if type(logger) == str and logger in logging.getLogger().manager.loggerDict:
+    logger = logging.getLogger(logger_name)
+    if type == "error": return logger.error(message)
+    if logger: logger.info(message)
+    if print_to_std: print(message + "\n")
+
+
+def expert_response_choice_or_question(messages, options_dict, self_consistency=1, **kwargs):
+    """
+    Implicit Abstain
+    """
+    log_info(f"++++++++++++++++++++ Start of Implicit Abstention [expert_basics.py:expert_response_choice_or_question()] ++++++++++++++++++++")
+    log_info(f"[<IMPLICIT ABSTAIN PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+    answers, questions, response_texts = [], [], {}
+    total_tokens = {"input_tokens": 0, "output_tokens": 0}
+    choice_logprobs = []
+    for i in range(self_consistency):
+        log_info(f"-------------------- Self-Consistency Iteration {i+1} --------------------")
+        response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+        total_tokens["input_tokens"] += num_tokens["input_tokens"]
+        total_tokens["output_tokens"] += num_tokens["output_tokens"]
+        if not response_text: 
+            log_info("[<IMPLICIT ABSTAIN LM RES>]: " + "No response --> Re-prompt")
+            continue
+        log_info("[<IMPLICIT ABSTAIN LM RES>]: " + response_text)
+        response_text = response_text.replace("Confident --> Answer: ", "").replace("Not confident --> Doctor Question: ", "")
+
+        if "?" not in response_text:
+            letter_choice = parse_choice(response_text, options_dict)
+            if letter_choice:
+                log_info("[<IMPLICIT ABSTAIN PARSED>]: " + letter_choice)
+                answers.append(letter_choice)
+                response_texts[letter_choice] = response_text
+                choice_logprobs.append(log_probs)
+        else:
+            # not a choice, parse as question
+            atomic_question = parse_atomic_question(response_text)
+            if atomic_question:
+                log_info("[<IMPLICIT ABSTAIN PARSED>]: " + atomic_question)
+                questions.append(atomic_question)
+                response_texts[atomic_question] = response_text
+            
+            else:
+                log_info("[<IMPLICIT ABSTAIN PARSED>]: " + "FAILED TO PARSE --> Re-prompt")
+
+    if len(answers) + len(questions) == 0:
+        log_info("[<IMPLICIT ABSTAIN SC-PARSED>]: " + "No response.")
+        return "No response.", None, None, 0.0, {}, total_tokens
+
+    conf_score = len(answers) / (len(answers) + len(questions))
+    if len(answers) > len(questions): 
+        final_answer = max(set(answers), key = answers.count)
+        response_text = response_texts[final_answer]
+        top_logprobs = choice_logprobs[answers.index(final_answer)]
+        atomic_question = None
+    else:
+        final_answer = None
+        rand_id = random.choice(range(len(questions)))
+        atomic_question = questions[rand_id]
+        response_text = response_texts[atomic_question]
+        top_logprobs = None
+    log_info(f"[<IMPLICIT ABSTAIN RETURN>]: atomic_question: {atomic_question}, final_answer: {final_answer}, conf_score: {conf_score} ([{len(answers)} : {len(questions)}])")
+    return response_text, atomic_question, final_answer, conf_score, top_logprobs, total_tokens
+
+
+
+def expert_response_yes_no(messages, self_consistency=1, **kwargs):
+    """
+    Binary Abstain
+    """
+    log_info(f"++++++++++++++++++++ Start of YES/NO Decision [expert_basics.py:expert_response_yes_no()] ++++++++++++++++++++")
+    log_info(f"[<YES/NO PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+
+    yes_no_responses, log_probs_list, response_texts = [], [], {}
+    total_tokens = {"input_tokens": 0, "output_tokens": 0}
+    for i in range(self_consistency):
+        log_info(f"-------------------- Self-Consistency Iteration {i+1} --------------------")
+        response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+        total_tokens["input_tokens"] += num_tokens["input_tokens"]
+        total_tokens["output_tokens"] += num_tokens["output_tokens"]
+        if not response_text: 
+            log_info("[<YES/NO LM RES>]: " + "No response.")
+        log_info("[<YES/NO LM RES>]: " + response_text)
+        log_probs_list.append(log_probs)
+
+        yes_choice = parse_yes_no(response_text)
+        log_info("[<YES/NO PARSED>]: " + yes_choice)
+        yes_no_responses.append(yes_choice)
+        response_texts[yes_choice] = response_text
+    
+    if yes_no_responses.count("YES") > yes_no_responses.count("NO"):
+        yes_choice = "YES"
+        log_probs = log_probs_list[yes_no_responses.index("YES")]
+    else:
+        yes_choice = "NO"
+        log_probs = log_probs_list[yes_no_responses.index("NO")]
+    log_info(f"[<YES/NO RETURN>]: yes_choice: {yes_choice}, confidence: {yes_no_responses.count('YES')/len(yes_no_responses)}")
+    return response_texts[yes_choice], yes_choice, yes_no_responses.count("YES")/len(yes_no_responses), log_probs, total_tokens
+
+
+
+def expert_response_confidence_score(messages, self_consistency=1, **kwargs):
+    """
+    Numerical Abstain
+    """
+    log_info(f"++++++++++++++++++++ Start of Numerical Confidence Score [expert_basics.py:expert_response_confidence_score()] ++++++++++++++++++++")
+    log_info(f"[<CONF SCORE PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+
+    conf_scores, log_probs_list, response_texts = [], {}, {}
+    total_tokens = {"input_tokens": 0, "output_tokens": 0}
+    for i in range(self_consistency):
+        log_info(f"-------------------- Self-Consistency Iteration {i+1} --------------------")
+        response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+        total_tokens["input_tokens"] += num_tokens["input_tokens"]
+        total_tokens["output_tokens"] += num_tokens["output_tokens"]
+        if not response_text: 
+            log_info("[<CONF SCORE LM RES>]: " + "No response.")
+            continue
+        log_info("[<CONF SCORE LM RES>]: " + response_text)
+
+        conf_score = parse_confidence_score(response_text)
+        conf_scores.append(conf_score)
+        log_probs_list[conf_score] = log_probs
+        response_texts[conf_score] = response_text
+        log_info(f"[<CONF SCORE PARSED>]: {conf_score}")
+    
+    if len(conf_scores) > 0:
+        avg_conf_score = sum(conf_scores) / len(conf_scores)
+        # response_text = "CONFIDENCE SCORE: " + str(avg_conf_score)
+        temp = [abs(r-avg_conf_score) for r in conf_scores]
+        response_text = response_texts[conf_scores[temp.index(min(temp))]]
+        log_probs = log_probs_list[conf_scores[temp.index(min(temp))]]
+    else:
+        avg_conf_score, response_text, log_probs = 0, "No response.", None
+    log_info(f"[<CONF SCORE RETURN>] (average conf score): {avg_conf_score}")
+    return response_text, avg_conf_score, log_probs, total_tokens
+
+
+
+def expert_response_scale_score(messages, self_consistency=1, **kwargs):
+    """
+    Scale Abstain
+    """
+    log_info(f"++++++++++++++++++++ Start of Scale Confidence Score [expert_basics.py:expert_response_scale_score()] ++++++++++++++++++++")
+    log_info(f"[<SCALE SCORE PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+
+    conf_scores, log_probs_list, response_texts = [], {}, {}
+    total_tokens = {"input_tokens": 0, "output_tokens": 0}
+    for i in range(self_consistency):
+        log_info(f"-------------------- Self-Consistency Iteration {i+1} --------------------")
+        response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+        total_tokens["input_tokens"] += num_tokens["input_tokens"]
+        total_tokens["output_tokens"] += num_tokens["output_tokens"]
+        if not response_text:
+            log_info("[<SCALE SCORE LM RES>]: " + "No response.")
+            continue
+        log_info("[<SCALE SCORE LM RES>]: " + response_text)
+
+        conf_score = parse_likert_scale(response_text)
+        conf_scores.append(conf_score)
+        log_probs_list[conf_score] = log_probs
+        response_texts[conf_score] = response_text
+        log_info("[<SCALE SCORE PARSED>]: " + str(conf_score))
+    
+    if len(conf_scores) > 0:
+        avg_conf_score = sum(conf_scores) / len(conf_scores)
+        temp = [abs(r-avg_conf_score) for r in conf_scores]
+        response_text = response_texts[conf_scores[temp.index(min(temp))]]
+        log_probs = log_probs_list[conf_scores[temp.index(min(temp))]]
+    else:
+        avg_conf_score, response_text, log_probs = 0, "No response.", None
+    log_info(f"[<SCALE SCORE RETURN>] (average conf score]): {avg_conf_score}")
+    return response_text, avg_conf_score, log_probs, total_tokens
+
+
+
+def expert_response_choice(messages, options_dict, **kwargs):
+    """
+    Get intermediate answer choice regardless of abstention decision
+    """
+    log_info(f"++++++++++++++++++++ Start of Multiple Chocie Decision [expert_basics.py:expert_response_choice()] ++++++++++++++++++++")
+    log_info(f"[<CHOICE PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+    response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+    if not response_text: 
+        log_info("[<CHOICE LM RES>]: " + "No response.")
+        return "No response.", None, num_tokens
+    log_info("[<CHOICE LM RES>]: " + response_text)
+
+    letter_choice = parse_choice(response_text, options_dict)
+    if letter_choice:
+        log_info("[<CHOICE PARSED>]: " + letter_choice)
+    else:
+        log_info("[<CHOICE PARSED>]: " + "FAILED TO PARSE.")
+    
+    return response_text, letter_choice, num_tokens
+
+
+
+def expert_response_question(messages, **kwargs):
+    """
+    Get follow-up question
+    """
+    log_info(f"++++++++++++++++++++ Start of Question Generator [expert_basics.py:expert_response_question()] ++++++++++++++++++++")
+    log_info(f"[<QUESTION GENERATOR PROMPT>] [len(messages)={len(messages)}] (messages[-1]):\n{messages[-1]['content']}")
+    response_text, log_probs, num_tokens = get_response(messages, **kwargs)
+    if not response_text: 
+        log_info("[<QUESTION GENERATOR LM RES>]: " + "No response.")
+        return "No response.", None, num_tokens
+    log_info("[<QUESTION GENERATOR LM RES>]: " + response_text)
+
+    atomic_question = parse_atomic_question(response_text)
+    if atomic_question:
+        log_info("[<QUESTION GENERATOR PARSED>]: " + atomic_question)
+    else:
+        log_info("[<QUESTION GENERATOR PARSED>]: " + "FAILED TO PARSE.")
+    
+    return response_text, atomic_question, num_tokens
+
+
+
+############################
+# Helper Functions for Parsing Responses
+############################
+
+def parse_atomic_question(response_text):
+    questions = []
+    for line in response_text.split("\n"):
+        if '?' in line:
+            questions.append(line.split(":")[-1].strip())
+        
+    if len(questions) == 0:
+        log_info("can't find question in answer: {}".format(response_text), type="error")
+        return None
+            
+    atomic_question = questions[-1].replace("'", "").replace('"', "").strip()
+    return atomic_question
+
+def parse_choice(response_text, options_dict):
+    if response_text.strip() in ["A", "B", "C", "D"]:
+        return response_text.strip()
+    for response_line in response_text.split("\n"):
+        for op_letter, op_text in options_dict.items():
+            if op_text.lower() in response_line.lower():
+                log_info(f"....Found {op_text} in response line: {response_line}")
+                return op_letter
+        for op_letter in options_dict.keys():
+            if op_letter in [token for token in re.sub(r"[,.;@#()?!'/&:$]+\ *", " ", response_line).split(' ')]:
+                # op_letter_str = str(op_letter) if op_letter else "none"
+                # response_line_str = str(response_line) if response_line else "none"
+                log_info(f"....Found {op_letter} in response line: {response_line}")
+                return op_letter
+    log_info("can't parse choice: {}".format(response_text), type="error")
+    return None
+
+def parse_yes_no(response_text):
+    temp_processed_response = response_text.lower().replace('.','').replace(',','').replace(';','').replace(':','').split("DECISION:")[-1].strip()
+    yes_answer = "yes" in temp_processed_response
+    no_answer = "no" in temp_processed_response
+    if yes_answer == no_answer:
+        yes_choice = "NO"
+        log_info("can't parse yes/no abstain answer: {}".format(response_text), type="error")
+    if yes_answer: yes_choice = "YES"
+    elif no_answer: yes_choice = "NO"
+    return yes_choice
+
+def parse_confidence_score(response_text):
+    # parse the probability
+    float_regex = re.compile(r'\d+\.\d+')
+    scores = re.findall(float_regex, response_text)
+
+    if len(scores) == 0:
+        log_info("can't parse confidence score - answer: {}".format(response_text), type="error")
+        score = round(0.2 + (random.random() - random.random()) * 0.2, 4)
+        return score
+    
+    prob = float(scores[-1])
+    if len(scores) > 1: logging.warning("more than one confidence score - using last: {}".format(response_text))
+    if prob > 1: logging.warning("confidence score > 1: {}".format(response_text))
+    return prob
+
+def parse_likert_scale(response_text):
+    temp_processed_response = response_text.lower().replace('.','').replace(',','').replace(';','').replace(':','')
+    if "very confident" in temp_processed_response:
+        conf_score = 5
+    elif "somewhat confident" in temp_processed_response:
+        conf_score = 4
+    elif "neither confident nor unconfident" in temp_processed_response:
+        conf_score = 3
+    elif "neither confident or unconfident" in temp_processed_response:
+        conf_score = 3
+    elif "somewhat unconfident" in temp_processed_response:
+        conf_score = 2
+    elif "very unconfident" in temp_processed_response:
+        conf_score = 1
+    else:
+        conf_score = 0
+        log_info("can't parse likert confidence score: {}".format(response_text), type="error")
+    return conf_score

src/expert_functions.py

@@ -0,0 +1,333 @@
+import prompts
+import expert_basics
+import logging
+
+PROB_THRESHOLD = 0.8
+SCALE_THRESHOLD = 4.0
+
+def answer_to_idx(answer):
+    return ord(answer) - ord("A")
+
+def log_info(message, logger="detail_logger", print_to_std=False):
+    if type(logger) == str and logger in logging.getLogger().manager.loggerDict:
+        logger = logging.getLogger(logger)
+    if logger: logger.info(message)
+    if print_to_std: print(message + "\n")
+
+
+
+def fixed_abstention_decision(max_depth, patient_state, inquiry, options_dict, **kwargs):
+    """
+    Fixed abstention strategy based on the current interaction length.
+    If the interaction length is less than the max depth, abstain, otherwise answer.
+    """
+    # first get the model's abstention decision
+    log_info(f"++++++++++++++++++++ Start of Fixed Abstention [expert_functions.py:fixed_abstention_decision()] ++++++++++++++++++++")
+    abstain_decision = len(patient_state['interaction_history']) < max_depth
+    conf_score = 1 if abstain_decision else 0
+    log_info(f"[ABSTENTION RESPONSE]: {abstain_decision}\n")
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+
+    prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+    messages_answer = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_answer}
+    ]
+    response_text, letter_choice, num_tokens = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+
+    log_info(f"[FIXED ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages_answer,
+        "letter_choice": letter_choice,
+    }
+
+
+
+def implicit_abstention_decision(patient_state, rationale_generation, inquiry, options_dict, **kwargs):
+    """
+    Implicit abstention strategy based on the current patient state.
+    This function uses the expert system to make a decision on whether to abstain or not based on the current patient state.
+    """
+    # Get the response from the expert system
+    prompt_key = "implicit_RG" if rationale_generation else "implicit"
+    abstain_task_prompt = prompts.expert_system[prompt_key]
+
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+    
+    # first get the model's abstention decision
+    prompt_abstain = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, abstain_task_prompt)
+
+    messages = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_abstain}
+    ]
+    response_text, atomic_question, letter_choice, conf_score, top_logprobs, num_tokens = expert_basics.expert_response_choice_or_question(messages, options_dict, **kwargs)
+    log_info(f"[ABSTENTION PROMPT]: {messages}")
+    log_info(f"[ABSTENTION RESPONSE]: {response_text}\n")
+    messages.append({"role": "assistant", "content": response_text})
+
+    if atomic_question != None: abstain_decision = True  # if the model generates a question, it is abstaining from answering, therefore abstain decision is True
+    elif letter_choice != None: abstain_decision = False  # if the model generates an answer, it is not abstaining from answering, therefore abstain decision is False
+    else: abstain_decision = True  # if the model generates neither an answer nor a question, it is abstaining from answering, therefore abstain decision is True
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    # note that we get this for free if implicit abstain already chooses an answer instead of a question
+    if letter_choice == None:
+        prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+        messages_answer = [
+            {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+            {"role": "user", "content": prompt_answer}
+        ]
+        response_text, letter_choice, num_tokens_answer = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+        num_tokens["input_tokens"] += num_tokens_answer["input_tokens"]
+        num_tokens["output_tokens"] += num_tokens_answer["output_tokens"]
+
+    log_info(f"[IMPLICIT ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}, atomic_question: {atomic_question}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages,
+        "letter_choice": letter_choice,
+        "atomic_question": atomic_question,
+    }
+
+
+
+def binary_abstention_decision(patient_state, rationale_generation, inquiry, options_dict, **kwargs):
+    """
+    Binary abstention strategy based on the current patient state.
+    This function prompts the user to make a binary decision on whether to abstain or not based on the current patient state.
+    """
+    # Get the response from the expert system
+    prompt_key = "binary_RG" if rationale_generation else "binary"
+    abstain_task_prompt = prompts.expert_system[prompt_key]
+
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+    
+    # first get the model's abstention decision
+    prompt_abstain = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, abstain_task_prompt)
+
+    messages = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_abstain}
+    ]
+    response_text, abstain_decision, conf_score, log_probs, num_tokens = expert_basics.expert_response_yes_no(messages, **kwargs)
+    abstain_decision = abstain_decision.lower() == 'no'
+    log_info(f"[ABSTENTION PROMPT]: {messages}")
+    log_info(f"[ABSTENTION RESPONSE]: {response_text}\n")
+    messages.append({"role": "assistant", "content": response_text})
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+    messages_answer = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_answer}
+    ]
+    response_text, letter_choice, num_tokens_answer = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+    num_tokens["input_tokens"] += num_tokens_answer["input_tokens"]
+    num_tokens["output_tokens"] += num_tokens_answer["output_tokens"]
+
+    log_info(f"[BINARY ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages,
+        "letter_choice": letter_choice,
+    }
+
+
+
+def numerical_abstention_decision(patient_state, rationale_generation, inquiry, options_dict, **kwargs):
+    """
+    Numerical abstention strategy based on the current patient state.
+    This function prompts the model to produce a numerical confidence score of how confident it is in its decision, then ask whether it wants to proceed
+    """
+
+    # Get the response from the expert system
+    prompt_key = "numerical_RG" if rationale_generation else "numerical"
+    abstain_task_prompt = prompts.expert_system[prompt_key]
+
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+    
+    # first get the model's abstention decision
+    prompt_abstain = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, abstain_task_prompt)
+
+    messages = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_abstain}
+    ]
+    response_text, conf_score, log_probs, num_tokens = expert_basics.expert_response_confidence_score(messages, **kwargs)
+    messages.append({"role": "assistant", "content": response_text})
+    
+    messages.append({"role": "user", "content": prompts.expert_system["yes_no"]})
+    # third return is supposed to be the conf_score in the binary setup, but we don't use it here because has conf score from last turn.
+    response_text, abstain_decision, _, log_probs, num_tokens_2 = expert_basics.expert_response_yes_no(messages, **kwargs)
+    abstain_decision = abstain_decision.lower() == 'no'
+    num_tokens["input_tokens"] += num_tokens_2["input_tokens"]
+    num_tokens["output_tokens"] += num_tokens_2["output_tokens"]
+    log_info(f"[ABSTENTION PROMPT]: {messages}")
+    log_info(f"[ABSTENTION RESPONSE]: {response_text}\n")
+    messages.append({"role": "assistant", "content": response_text})
+
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+    messages_answer = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_answer}
+    ]
+    response_text, letter_choice, num_tokens_answer = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+    num_tokens["input_tokens"] += num_tokens_answer["input_tokens"]
+    num_tokens["output_tokens"] += num_tokens_answer["output_tokens"]
+
+    log_info(f"[NUMERICAL ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages,
+        "letter_choice": letter_choice,
+    }
+
+
+
+def numcutoff_abstention_decision(patient_state, rationale_generation, inquiry, options_dict, abstain_threshold, **kwargs):
+    """
+    Numcutoff abstention strategy based on the current patient state.
+    This function prompts the model to produce a numerical confidence score of how confident it is in its decision, then decide abstention based on arbitrarily set threshold
+    """
+    if not abstain_threshold: abstain_threshold = PROB_THRESHOLD
+    
+    # Get the response from the expert system
+    prompt_key = "numcutoff_RG" if rationale_generation else "numcutoff"
+    abstain_task_prompt = prompts.expert_system[prompt_key]
+
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+    
+    # first get the model's abstention decision
+    prompt_abstain = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, abstain_task_prompt)
+
+    messages = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_abstain}
+    ]
+    response_text, conf_score, log_probs, num_tokens = expert_basics.expert_response_confidence_score(messages, abstain_threshold=abstain_threshold, **kwargs)
+    abstain_decision = conf_score < abstain_threshold
+    log_info(f"[ABSTENTION PROMPT]: {messages}")
+    log_info(f"[ABSTENTION RESPONSE]: {response_text}\n")
+    messages.append({"role": "assistant", "content": response_text})
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+    messages_answer = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_answer}
+    ]
+    response_text, letter_choice, num_tokens_answer = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+    num_tokens["input_tokens"] += num_tokens_answer["input_tokens"]
+    num_tokens["output_tokens"] += num_tokens_answer["output_tokens"]
+
+    log_info(f"[NUMCUTOFF ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages,
+        "letter_choice": letter_choice,
+    }
+
+
+
+def scale_abstention_decision(patient_state, rationale_generation, inquiry, options_dict, abstain_threshold, **kwargs):
+    """
+    Likert abstention strategy based on the current patient state.
+    This function prompts the model to produce a likert scale confidence score of how confident it is in its decision, then decide abstention based on a cutoff
+    """
+    if not abstain_threshold: abstain_threshold = SCALE_THRESHOLD
+
+    # Get the response from the expert system
+    prompt_key = "scale_RG" if rationale_generation else "scale"
+    abstain_task_prompt = prompts.expert_system[prompt_key]
+
+    patient_info = patient_state["initial_info"]
+    conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+    options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+    
+    # first get the model's abstention decision
+    prompt_abstain = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, abstain_task_prompt)
+
+    messages = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_abstain}
+    ]
+    response_text, conf_score, log_probs, num_tokens = expert_basics.expert_response_scale_score(messages, abstain_threshold=abstain_threshold, **kwargs)
+    abstain_decision = conf_score < abstain_threshold
+    log_info(f"[ABSTENTION PROMPT]: {messages}")
+    log_info(f"[ABSTENTION RESPONSE]: {response_text}\n")
+    messages.append({"role": "assistant", "content": response_text})
+
+    # second, no matter what the model's abstention decision is, get an intermediate answer for evaluation and analysis
+    prompt_answer = prompts.expert_system["curr_template"].format(patient_info, conv_log if conv_log != '' else 'None', inquiry, options_text, prompts.expert_system["answer"])
+    messages_answer = [
+        {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+        {"role": "user", "content": prompt_answer}
+    ]
+    response_text, letter_choice, num_tokens_answer = expert_basics.expert_response_choice(messages_answer, options_dict, **kwargs)
+    num_tokens["input_tokens"] += num_tokens_answer["input_tokens"]
+    num_tokens["output_tokens"] += num_tokens_answer["output_tokens"]
+
+    log_info(f"[SCALE ABSTAIN RETURN]: abstain: {abstain_decision}, confidence: {conf_score}, letter_choice: {letter_choice}, usage: {num_tokens}\n")
+    return {
+        "abstain": abstain_decision,
+        "confidence": conf_score,
+        "usage": num_tokens,
+        "messages": messages,
+        "letter_choice": letter_choice,
+    }
+
+
+
+def question_generation(patient_state, inquiry, options_dict, messages, independent_modules, **kwargs):
+    task_prompt = prompts.expert_system["atomic_question_improved"]
+
+    if independent_modules:
+        patient_info = patient_state["initial_info"]
+        conv_log = '\n'.join([f"{prompts.expert_system['question_word']}: {qa['question']}\n{prompts.expert_system['answer_word']}: {qa['answer']}" for qa in patient_state["interaction_history"]])
+        options_text = f'A: {options_dict["A"]}, B: {options_dict["B"]}, C: {options_dict["C"]}, D: {options_dict["D"]}'
+        prompt = prompts.expert_system["curr_template"].format(patient_info, conv_log, inquiry, options_text, task_prompt)
+
+        messages = [
+            {"role": "system", "content": prompts.expert_system["meditron_system_msg"]},
+            {"role": "user", "content": prompt}
+        ]
+    else:
+        messages.append({"role": "user", "content": task_prompt})
+
+    response_text, atomic_question, num_tokens = expert_basics.expert_response_question(messages, **kwargs)
+    log_info(f"[ATOMIC QUESTION PROMPT]: {messages}")
+    log_info(f"[ATOMIC QUESTION RESPONSE]: {atomic_question}\n")
+    messages.append({"role": "assistant", "content": atomic_question})
+
+    log_info(f"[ATOMIC QUESTION RETURN]: {atomic_question}, usage: {num_tokens}\n")
+    return {
+        "atomic_question": atomic_question,
+        "messages": messages,
+        "usage": num_tokens,
+    }

src/helper.py

@@ -0,0 +1,158 @@
+import torch
+import logging
+from keys import mykey
+
+# A dictionary to cache models and tokenizers to avoid reloading
+
+global models
+models = {}
+
+def log_info(message, logger_name="message_logger", print_to_std=False, mode="info"):
+    logger = logging.getLogger(logger_name)
+    if logger: 
+        if mode == "error": logger.error(message)
+        if mode == "warning": logger.warning(message)
+        else: logger.info(message)
+    if print_to_std: print(message + "\n")
+
+class ModelCache:
+    def __init__(self, model_name, use_vllm=False, use_api=None, **kwargs):
+        self.model_name = model_name
+        self.use_vllm = use_vllm
+        self.use_api = use_api
+        self.model = None
+        self.tokenizer = None
+        self.terminators = None
+        self.client = None
+        self.args = kwargs
+        self.load_model_and_tokenizer()
+    
+    def load_model_and_tokenizer(self):
+        if self.use_api == "openai":
+            from openai import OpenAI
+            self.api_account = self.args.get("api_account", "openai")
+            self.client = OpenAI(api_key=mykey[self.api_account]) # Setup API key appropriately in keys.py
+        elif self.use_vllm:
+            try:
+                from vllm import LLM
+                enable_prefix_caching = self.args.get("enable_prefix_caching", False)
+                self.model = LLM(model=self.model_name, enable_prefix_caching=enable_prefix_caching)
+                from transformers import AutoTokenizer
+                self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
+                self.tokenizer.pad_token = self.tokenizer.eos_token
+                self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+                self.terminators = [self.tokenizer.eos_token_id, self.tokenizer.convert_tokens_to_ids("<|eot_id|>")]
+            except Exception as e:
+                log_info(f"[ERROR] [{self.model_name}]: If using a custom local model, it is not compatible with VLLM, will load using Huggingfcae and you can ignore this error: {str(e)}", mode="error")
+                self.use_vllm = False
+        if not self.use_vllm and self.use_api != "openai":
+            from transformers import AutoModelForCausalLM, AutoTokenizer
+            self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
+            self.model = AutoModelForCausalLM.from_pretrained(self.model_name)
+            self.model.eval()  # Set the model to evaluation mode
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+            self.terminators = [self.tokenizer.eos_token_id, self.tokenizer.convert_tokens_to_ids("<|eot_id|>")]
+    
+    def generate(self, messages):
+        log_info(f"[{self.model_name}][INPUT]: {messages}")
+
+        self.temperature = self.args.get("temperature", 0.6)
+        self.max_tokens = self.args.get("max_tokens", 256)
+        self.top_p = self.args.get("top_p", 0.9)
+        self.top_logprobs = self.args.get("top_logprobs", 0)
+
+        if self.use_api == "openai": self.openai_generate(messages)
+        elif self.use_vllm: return self.vllm_generate(messages)
+        else: return self.huggingface_generate(messages)
+    
+    def huggingface_generate(self, messages):
+        try:
+            inputs = self.tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(self.model.device)
+        except:
+            # Join messages into a single prompt for general language models
+            log_info(f"[{self.model_name}]: Could not apply chat template to messages.", mode="warning")
+            prompt = "\n\n".join([m['content'] for m in messages])
+            inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
+
+        outputs = self.model.generate(
+            inputs,
+            do_sample=True,
+            max_new_tokens=self.max_tokens, 
+            temperature=self.temperature,
+            top_p=self.top_p,
+            pad_token_id=self.tokenizer.pad_token_id,
+            eos_token_id=self.terminators
+        )
+        # TODO: If top_logprobs > 0, return logprobs of generation
+        response_text = self.tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
+        usage = {"input_tokens": inputs.shape[-1], "output_tokens": outputs.shape[-1]-inputs.shape[-1]}
+        output_dict = {'response_text': response_text, 'usage': usage}
+
+        log_info(f"[{self.model_name}][OUTPUT]: {output_dict}")
+        return response_text, None, usage
+        
+    def vllm_generate(self, messages):
+        try:
+            inputs = self.tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+        except:
+            # Join messages into a single prompt for general language models
+            log_info(f"[{self.model_name}]: Could not apply chat template to messages.", mode="warning")
+            inputs = "\n\n".join([m['content'] for m in messages])
+            # inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
+
+        from vllm import SamplingParams
+        frequency_penalty = self.args.get("frequency_penalty", 0)
+        presence_penalty = self.args.get("presense_penalty", 0)
+        sampling_params = SamplingParams(temperature=self.temperature, max_tokens=self.max_tokens, top_p=self.top_p, logprobs=self.top_logprobs, 
+                                        frequency_penalty=frequency_penalty, presence_penalty=presence_penalty)
+        
+        outputs = self.model.generate(inputs, sampling_params)
+        response_text = outputs[0].outputs[0].text
+        logprobs = outputs[0].outputs[0].cumulative_logprob
+        # TODO: If top_logprobs > 0, return logprobs of generation
+        # if self.top_logprobs > 0: logprobs = outputs[0].outputs[0].logprobs
+        usage = {"input_tokens": len(outputs[0].prompt_token_ids), "output_tokens": len(outputs[0].outputs[0].token_ids)}
+        output_dict = {'response_text': response_text, 'usage': usage}
+
+        log_info(f"[{self.model_name}][OUTPUT]: {output_dict}")
+        return response_text, logprobs, usage
+
+    def openai_generate(self, messages):
+        if self.top_logprobs == 0:
+            response = self.client.chat.completions.create(
+                        model=self.model_name,
+                        messages=messages,
+                        temperature=self.temperature,
+                        max_tokens=self.max_tokens,
+                        top_p=self.top_p
+                    )
+        else:
+            response = self.client.chat.completions.create(
+                        model=self.model_name,
+                        messages=messages,
+                        temperature=self.temperature,
+                        max_tokens=self.max_tokens,
+                        top_p=self.top_p,
+                        logprobs=True, 
+                        top_logprobs=self.top_logprobs
+                    )
+        
+        num_input_tokens = response["usage"]["prompt_tokens"]
+        num_output_tokens = response["usage"]["completion_tokens"]
+        response_text = response.choices[0].text.strip()
+        log_probs = response.choices[0].logprobs.top_logprobs if self.top_logprobs > 0 else None
+        
+        log_info(f"[{self.model_name}][OUTPUT]: {response}")
+        return response_text, log_probs, {"input_tokens": num_input_tokens, "output_tokens": num_output_tokens}
+
+
+def get_response(messages, model_name, use_vllm=False, use_api=None, **kwargs):
+    if 'gpt' in model_name or 'o1' in model_name: use_api = "openai"
+    
+    model_cache = models.get(model_name, None)
+    if model_cache is None:
+        model_cache = ModelCache(model_name, use_vllm=use_vllm, use_api=use_api, **kwargs)
+        models[model_name] = model_cache
+    
+    return model_cache.generate(messages)

src/keys.py

@@ -0,0 +1,3 @@
+mykey = {
+    "mediQ": "sk-1234567890abcdef1234567890abcdef12345678",
+}

src/mediQ_benchmark.py

@@ -0,0 +1,177 @@
+import json
+import os
+import time
+import logging
+from args import get_args
+from patient import Patient
+import importlib
+
+def setup_logger(name, file):
+    if not file: return None
+    logger = logging.getLogger(name)
+    handler = logging.FileHandler(file, mode='a')
+    formatter = logging.Formatter('[%(asctime)s] [%(levelname)s] %(message)s')
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+    logger.setLevel(logging.INFO)
+    return logger
+
+def log_info(message, print_to_std=False):
+    if history_logger: history_logger.info(message)
+    if detail_logger: detail_logger.info(message)
+    if print_to_std: print(message + "\n")
+
+def load_data(filename):
+    with open(filename, "r") as json_file:
+        json_list = list(json_file)
+    data = [json.loads(line) for line in json_list]
+    data = {item['id']: item for item in data}
+    return data
+
+def main():
+    if os.path.exists(args.output_filename):
+        with open(args.output_filename, "r") as f:
+            lines = f.readlines()
+        output_data = [json.loads(line) for line in lines]
+        if len(lines) == 0: processed_ids = []
+        else: processed_ids = {sample["id"]: {"correct": sample["interactive_system"]["letter_choice"] == sample["info"]["correct_answer_idx"],
+                                              "timeout": len(sample["interactive_system"]["intermediate_choices"]) > args.max_questions,
+                                              "turns": sample["interactive_system"]["num_questions"]}
+                                for sample in output_data}
+    else:
+        processed_ids = []
+
+    expert_module = importlib.import_module(args.expert_module)
+    expert_class = getattr(expert_module, args.expert_class)
+    patient_module = importlib.import_module(args.patient_module)
+    patient_class = getattr(patient_module, args.patient_class)
+    
+    patient_data_path = os.path.join(args.data_dir, args.dev_filename)
+    patient_data = load_data(patient_data_path)
+
+    num_processed = 0
+    correct_history, timeout_history, turn_lengths = [], [], []
+
+    for pid, sample in patient_data.items():
+        if pid in processed_ids:
+            print(f"Skipping patient {pid} as it has already been processed.")
+            correct_history.append(processed_ids[pid]["correct"])
+            timeout_history.append(processed_ids[pid]["timeout"])
+            turn_lengths.append(processed_ids[pid]["turns"])
+            continue
+
+        log_info(f"|||||||||||||||||||| PATIENT #{pid} ||||||||||||||||||||")
+        letter_choice, questions, answers, temp_choice_list, temp_additional_info, sample_info = run_patient_interaction(expert_class, patient_class, sample)
+        log_info(f"|||||||||||||||||||| Interaction ended for patient #{pid} ||||||||||||||||||||\n\n\n")
+
+        output_dict = {
+            "id": pid,
+            "interactive_system": {
+                "correct": letter_choice == sample["answer_idx"],
+                "letter_choice": letter_choice,
+                "questions": questions,
+                "answers": answers,
+                "num_questions": len(questions),
+                "intermediate_choices": temp_choice_list,
+                "temp_additional_info": temp_additional_info
+            },
+            "info": sample_info,
+            # TODO: add additional evaluation metrics for analysis, some metrics can be found in src/evaluate.py
+            # "eval": {
+            #     "confidence_scores": [],
+            #     "repeat_question_score": [],
+            #     "repeat_answer_score": [],
+            #     "relevancy_score": [],
+            #     "delta_confidence_score": [],
+            #     "specificity_score": []
+            # }
+        }
+
+        # create the directory if it does not exist
+        os.makedirs(os.path.dirname(args.output_filename), exist_ok=True)
+        with open(args.output_filename, 'a+') as f:
+            f.write(json.dumps(output_dict) + '\n')
+
+        correct_history.append(letter_choice == sample["answer_idx"])
+        timeout_history.append(len(temp_choice_list) > args.max_questions)
+        turn_lengths.append(len(temp_choice_list))
+        num_processed += 1
+        accuracy = sum(correct_history) / len(correct_history) if len(correct_history) > 0 else None
+        timeout_rate = sum(timeout_history) / len(timeout_history) if len(timeout_history) > 0 else None
+        avg_turns = sum(turn_lengths) / len(turn_lengths) if len(turn_lengths) > 0 else None
+
+        results_logger.info(f'Processed {num_processed}/{len(patient_data)} patients | Accuracy: {accuracy}')
+        print(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] Processed {num_processed}/{len(patient_data)} patients | Accuracy: {accuracy} | Timeout Rate: {timeout_rate} | Avg. Turns: {avg_turns}")
+    print(f"Accuracy: {sum(correct_history)} / {len(correct_history)} = {accuracy}")
+    print(f"Timeout Rate: {sum(timeout_history)} / {len(timeout_history)} = {timeout_rate}")
+    print(f"Avg. Turns: {avg_turns}")
+    
+
+def run_patient_interaction(expert_class, patient_class, sample):
+    expert_system = expert_class(args, sample["question"], sample["options"])
+    patient_system = patient_class(args, sample)  # Assuming the patient_system is initialized with the sample which includes necessary context
+    temp_choice_list = []
+    temp_additional_info = []  # To store optional data like confidence scores
+
+    while len(patient_system.get_questions()) < args.max_questions:
+        log_info(f"==================== Turn {len(patient_system.get_questions()) + 1} ====================")
+        patient_state = patient_system.get_state()
+        response_dict = expert_system.respond(patient_state)
+        log_info(f"[Expert System]: {response_dict}")
+        
+        # Optional return values for analysis, e.g., confidence score, logprobs
+        temp_additional_info.append({k: v for k, v in response_dict.items() if k not in ["type", "letter_choice", "question"]})
+
+        if response_dict["type"] == "question":
+            # still make the Expert generate a choice based on the current state for intermediate evaluation, log the question as an intermediate choice
+            temp_choice_list.append(response_dict["letter_choice"])
+            # Patient generates an answer based on the last question asked, and add to memory
+            patient_response = patient_system.respond(response_dict["question"])
+            log_info(f"[Patient System]: {patient_response}")
+
+        elif response_dict["type"] == "choice":
+            expert_decision = response_dict["letter_choice"]
+            temp_choice_list.append(expert_decision)
+            sample_info = {
+                "initial_info": patient_system.initial_info,
+                "correct_answer": sample["answer"],
+                "correct_answer_idx": sample["answer_idx"],
+                "question": sample["question"],
+                "options": sample["options"],
+                "context": sample["context"],
+                "facts": patient_system.facts, # if the FactSelectPatient patient module is used, this will store the atomic facts the patient used to answer questions for reproducibility
+            }
+            return expert_decision, patient_system.get_questions(), patient_system.get_answers(), temp_choice_list, temp_additional_info, sample_info
+        
+        else:
+            raise ValueError("Invalid response type from expert_system.")
+    
+    # If max questions are reached and no final decision has been made
+    log_info(f"==================== Max Interaction Length ({args.max_questions} turns) Reached --> Force Final Answer ====================")
+    patient_state = patient_system.get_state()
+    response_dict = expert_system.respond(patient_state)
+    log_info(f"[Expert System]: {response_dict}")
+    stuck_response = response_dict["letter_choice"]
+    # Optional return values for analysis, e.g., confidence score, logprobs
+    temp_additional_info.append({k: v for k, v in response_dict.items() if k != "letter_choice"})
+
+    sample_info = {
+        "initial_info": patient_system.initial_info,
+        "correct_answer": sample["answer"],
+        "correct_answer_idx": sample["answer_idx"],
+        "question": sample["question"],
+        "options": sample["options"],
+        "context": sample["context"],
+        "facts": patient_system.facts, # if the FactSelectPatient patient module is used, this will store the atomic facts the patient used to answer questions for reproducibility
+    }
+    
+    return stuck_response, patient_system.get_questions(), patient_system.get_answers(), temp_choice_list + [stuck_response], temp_additional_info, sample_info
+
+
+if __name__ == "__main__":
+    args = get_args()
+    results_logger = setup_logger('results_logger', args.log_filename)
+    history_logger = setup_logger('history_logger', args.history_log_filename)
+    detail_logger = setup_logger('detail_logger', args.detail_log_filename)
+    message_logger = setup_logger('message_logger', args.message_log_filename)
+    main()

src/patient.py

@@ -0,0 +1,105 @@
+import random
+from helper import get_response
+
+class Patient:
+    def __init__(self, args, sample):
+        # Assuming 'context' is a list or a long string of historical or background information
+        if isinstance(sample['context'], list) and len(sample['context']) > 0:
+            if 'initial_info' in sample: self.initial_info = sample['initial_info']
+            else: self.initial_info = sample['context'][0]  # Taking the first item if it's a list
+            self.context_list = sample['context']
+            self.context_para = " ".join(sample['context'])
+        elif isinstance(sample['context'], str):
+            # Assuming sentences are separated by periods, taking the first sentence
+            if 'initial_info' in sample: self.initial_info = sample['initial_info']
+            else: self.initial_info = sample['context'].split(". ")[0]
+            temp = sample['context'].split(". ")
+            self.context_list = [temp[i]+'.' if i!=len(temp)-1 and not temp[i].endswith('.') else temp[i] for i in range(len(temp))]
+            self.context_para = sample['context']
+        else:
+            if 'initial_info' in sample: self.initial_info = sample['initial_info']
+            else: self.initial_info = ""  # Default fallback
+            self.context_list = []
+            self.context_para = 'None'
+        
+        self.model_name = args.patient_model
+        self.history = []  # To track the interaction history of questions and answers
+        self.facts = sample['atomic_facts'] if 'atomic_facts' in sample else None  # To store atomic facts after initial processing, you can choose to store this somewhere locally to avoid repeated processing
+
+        self.max_length = 50  # Maximum length of the response (different from the expert system)
+        self.use_vllm = args.use_vllm  
+        self.use_api = args.use_api  # Use an API to generate responses
+
+    def update_state(self, question, answer):
+        # Update the internal history with the new question and the corresponding answer
+        self.history.append({"question": question, "answer": answer})
+
+    def get_state(self):
+        # Return the initial context and the history of interactions
+        return {
+            "initial_info": self.initial_info,
+            "interaction_history": self.history
+        }
+    
+    def get_questions(self):
+        # Return the list of questions asked so far
+        return [qa["question"] for qa in self.history]
+    
+    def get_answers(self):
+        # Return the list of answers provided so far
+        return [qa["answer"] for qa in self.history]
+    
+    def get_response(self, messages, max_length=None):
+        if max_length is None: max_length = self.max_length
+        return get_response(messages, self.model_name, use_vllm=self.use_vllm, use_api=self.use_api, max_length=max_length)
+    
+    def respond(self, question):
+        raise NotImplementedError
+    
+
+class RandomPatient(Patient):
+    def respond(self, question):
+        # Randomly select a response mode
+        if random.random() < 0.5 or len(self.context_list) == 0:
+            answer = "The patient cannot answer this question, please do not ask this question again."
+        else:
+            answer = random.choice(self.context_list)
+        self.update_state(question, answer)
+        return answer
+
+class DirectPatient(Patient):
+    def respond(self, question):
+        system_prompt = "Answer the question with the given context."
+        user_prompt = f"Context: \"{self.initial_info}\"\nQuestion: \"{question}\"\n"
+        messages = [{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}]
+        response, log_probs, num_tokens = self.get_response(messages)
+        self.update_state(question, response)
+        return response
+
+class InstructPatient(Patient):
+    def respond(self, question):
+        system_prompt = "You are a truthful assistant that understands the patient's information, and you are trying to answer questions from a medical doctor about the patient."
+        user_prompt = f"Below is a context paragraph describing the patient and their conditions:\n\"{self.context_para}\"\nQuestion from the doctor: \"{question}\"\nUse the context paragraph to answer the doctor's question. If the paragraph does not answer the question, simply say \"The patient cannot answer this question, please do not ask this question again.\" Answer only what the question asks for. Do not provide any analysis, inference, or implications. Respond with a straightforward answer to the question ONLY and NOTHING ELSE."
+        messages = [{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}]
+        response, log_probs, num_tokens = self.get_response(messages)
+        self.update_state(question, response)
+        return response
+    
+class FactSelectPatient(Patient):
+    def respond(self, question):
+        if not self.facts:
+            # Decompose context into facts if not already done
+            system_prompt = "You are a truthful medical assistant that understands the patient's information."
+            user_prompt = f"Break the following patient information into a list of independent atomic facts, with one piece of information in each statement. Each fact should only include the smallest unit of information, but should be self-contained.\n\"{self.context_para}\"\nResponse with the list of atomic facts and nothing else, prepend each fact by an index starting from 1. No sub-list allowed."
+            messages = [{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}]
+            response_text, log_probs, num_tokens = self.get_response(messages, max_length=1000)
+            response_text = [s.strip() for s in response_text.splitlines()]
+            self.facts = response_text
+        
+        facts_prompt = "\n".join(self.facts)
+        system_prompt = "You are a truthful medical assistant that understands the patient's information, and you are trying to answer questions from a medical doctor about the patient given a list of factual statements describing the patient. Please return the facts that answer the doctor's question verbatim without any additional information. If none of the facts answer the question, simply say \"The patient cannot answer this question, please do not ask this question again.\""
+        prompt = f"List of facts:\n{facts_prompt}\n\nQuestion from the doctor: \"{question}\"\n\nStatements that answer the question:"
+        messages = [{"role": "system", "content": system_prompt}, {"role": "user", "content": prompt}]
+        response, log_probs, num_tokens = self.get_response(messages)
+        self.update_state(question, response)
+        return response

src/prompts.py

@@ -0,0 +1,113 @@
+expert_system = {
+    "meditron_system_msg_old": "You are a medical doctor answering real-world medical entrance exam questions. Based on your understanding of basic and clinical science, medical knowledge, and mechanisms underlying health, disease, patient care, and modes of therapy, answer the following multiplechoice question. Base your answer on the current and standard practices referenced in medical guidelines.\nTask: You will be asked to reason through the current patient's information and either ask an information seeking question or choose an option.",
+
+    "meditron_system_msg_original": "You are a medical doctor answering real-world medical entrance exam questions. Based on your understanding of basic and clinical science, medical knowledge, and mechanisms underlying health, disease, patient care, and modes of therapy, answer the following multiple choice question. Base your answer on the current and standard practices referenced in medical guidelines.",
+
+    "meditron_system_msg": "You are a medical doctor trying to reason through a real-life clinical case. Based on your understanding of basic and clinical science, medical knowledge, and mechanisms underlying health, disease, patient care, and modes of therapy, respond according to the task specified by the user. Base your response on the current and standard practices referenced in medical guidelines.",
+
+    "basic_system_msg": "You are an experienced doctor trying to make a medical decision about a patient.",
+
+    "empty_system_msg": "",
+
+    "only_choice": "Please answer with ONLY the correct letter choice (JUST ONE LETTER and NOTHING ELSE): A, B, C, or D.",
+
+    "system": "You are an experienced doctor trying to make a medical decision about a patient.",
+    
+    "starter": """A patient comes into the clinic presenting with a symptom as described in the conversation log below:\n\nCONVERSATION LOG:\n""",
+
+    "question_word": "Doctor Question",
+    "answer_word": "Patient Response",
+
+    "task": "Given the information from above, your task is to choose one of four options that best answers the inquiry.",
+    
+    "prompt": """\nMedical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. How confident are you to pick the correct option to the inquiry factually using the conversation log? In the first line of your response, generate the probability as a float from 0 to 1.\n\nIf there are missing features that prevent you from picking a confident and factual answer to the inquiry, consider which features are not yet asked about in the conversation log; then, consider which missing feature is the most important to ask the patient in order to provide the most helpful information toward a correct medical decision. Ask ONE SPECIFIC ATOMIC QUESTION to address this feature. The question should be bite-sized, and NOT ask for too much at once. In the second line of your response, generate the atomic question and nothing else.\n\nHowever, if you feel like you already have enough information from the above question-answer pairs to answer the patient inquiry, use the above information to produce a factual conclusion. In this case, answer with ONLY the correct letter choice and nothing else.""",
+    
+    "yes_no": "Now, are you confident to pick the correct option to the inquiry factually using the conversation log? Answer with YES or NO and NOTHING ELSE.",
+
+    
+    "implicit": "Given the information so far, if you are confident to pick an option correctly and factually, respond with the letter choice and NOTHING ELSE. Otherwise, if you are not confident to pick an option and need more information, ask ONE SPECIFIC ATOMIC QUESTION to the patient. The question should be bite-sized, NOT ask for too much at once, and NOT repeat what has already been asked. In this case, respond with the atomic question and NOTHING ELSE.",
+
+    "implicit_RG": "Given the information so far, if you are confident to pick an option correctly and factually, respond in the format:\nREASON: a one-sentence explanation of why you are choosing a particular option.\nANSWER: the letter choice and NOTHING ELSE. Otherwise, if you are not confident to pick an option and need more information, ask ONE SPECIFIC ATOMIC QUESTION to the patient. The question should be bite-sized, NOT ask for too much at once, and NOT repeat what has already been asked. In this case, respond in the format:\nREASON: a one-sentence explanation of why you should ask the particular question.\nQUESTION: the atomic question and NOTHING ELSE.",
+
+    "binary": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. Now, are you confident to pick the correct option to the inquiry factually using the conversation log? Answer with YES or NO and NOTHING ELSE.",
+    
+    "binary_RG": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. Up to this point, are you confident to pick the correct option to the inquiry factually using the conversation log? Answer in the following format:\nREASON: a one-sentence explanation of why you are or are not confident and what other information is needed.\nDECISION: YES or NO.",
+
+    "numcutoff": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. What is your confidence score to pick the correct option to the inquiry factually using the conversation log? Answer with the probability as a float from 0.0 to 1.0 and NOTHING ELSE.",
+
+    "numcutoff_RG": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. What is your confidence score to pick the correct option to the inquiry factually using the conversation log? Answer strictly in the following format:\nREASON: a one-sentence explanation of why you are or are not confident and what other information is needed.\nSCORE: your confidence score written as a float from 0.0 to 1.0.",
+
+    "numerical": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. What is your confidence score to pick the correct option to the inquiry factually using the conversation log? Answer with the probability as a float from 0.0 to 1.0 and NOTHING ELSE.",
+
+    "numerical_RG": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. What is your confidence score to pick the correct option to the inquiry factually using the conversation log? Answer strictly in the following format:\nREASON: a one-sentence explanation of why you are or are not confident and what other information is needed.\nSCORE: your confidence score written as a float from 0.0 to 1.0.",
+
+    "scale": """Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. How confident are you to pick the correct option to the problem factually using the conversation log? Choose between the following ratings:
+"Very Confident" - The correct option is supported by all evidence, and there is enough evidence to eliminate the rest of the answers, so the option can be confirmed conclusively.
+"Somewhat Confident" - I have reasonably enough information to tell that the correct option is more likely than other options, more information is helpful to make a conclusive decision.
+"Neither Confident or Unconfident" - There are evident supporting the correct option, but further evidence is needed to be sure which one is the correct option.
+"Somewhat Unconfident" - There are evidence supporting more than one options, therefore more questions are needed to further distinguish the options.
+"Very Unconfident" - There are not enough evidence supporting any of the options, the likelihood of picking the correct option at this point is near random guessing.\n\nThink carefully step by step, respond with the chosen confidence rating ONLY and NOTHING ELSE.""",
+
+    "scale_RG": """Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. How confident are you to pick the correct option to the problem factually using the conversation log? Choose between the following ratings:
+"Very Confident" - The correct option is supported by all evidence, and there is enough evidence to eliminate the rest of the answers, so the option can be confirmed conclusively.
+"Somewhat Confident" - I have reasonably enough information to tell that the correct option is more likely than other options, more information is helpful to make a conclusive decision.
+"Neither Confident or Unconfident" - There are evident supporting the correct option, but further evidence is needed to be sure which one is the correct option.
+"Somewhat Unconfident" - There are evidence supporting more than one options, therefore more questions are needed to further distinguish the options.
+"Very Unconfident" - There are not enough evidence supporting any of the options, the likelihood of picking the correct option at this point is near random guessing.\n\nAnswer in the following format:\nREASON: a one-sentence explanation of why you are or are not confident and what other information is needed.\nDECISION: chosen rating from the above list.""",
+
+    "yes_no": "Now, are you confident to pick the correct option to the inquiry factually using the conversation log? Answer with YES or NO and NOTHING ELSE.",
+
+    "verbal_abstain_llama": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. Up to this point, are you confident to pick the correct option to the inquiry factually using the conversation log? Answer in the following format:\nDECISION: YES or NO.",
+
+    "implicit_abstain": "Medical conditions are complex, so you should seek to understand their situations across many features. First, consider which medical specialty is this patient's case; then, consider a list of necessary features a doctor would need to make the right medical judgment; finally, consider whether all necessary information is given in the conversation above. In the following cases, either answer the question or ask another information-seeking question:\n1. If you are confident to pick the correct option to the inquiry factually using the conversation log, answer with ONLY the correct letter choice and NOTHING ELSE.\n2. If you are not confident to pick the correct option to the inquiry factually using the conversation log, consider what are the missing information that would help you differenciate among the options. Ask ONE SPECIFIC ATOMIC QUESTION to address the missing feature. The question should be bite-sized, and NOT ask for too much at once. Make sure to NOT repeat any questions from the above conversation log. Generate the atomic question and NOTHING ELSE.",
+ 
+    "atomic_question": "If there are missing features that prevent you from picking a confident and factual answer to the inquiry, consider which features are not yet asked about in the conversation log; then, consider which missing feature is the most important to ask the patient in order to provide the most helpful information toward a correct medical decision. Ask ONE SPECIFIC ATOMIC QUESTION to address this feature. The question should be bite-sized, and NOT ask for too much at once. Generate the atomic question and NOTHING ELSE.",
+    
+    "atomic_question_improved": "If there are missing features that prevent you from picking a confident and factual answer to the inquiry, consider which features are not yet asked about in the conversation log; then, consider which missing feature is the most important to ask the patient in order to provide the most helpful information toward a correct medical decision. You can ask about any relevant information about the patient’s case, such as family history, tests and exams results, treatments already done, etc. Consider what are the common questions asked in the specific subject relating to the patient’s known symptoms, and what the best and most intuitive doctor would ask. Ask ONE SPECIFIC ATOMIC QUESTION to address this feature. The question should be bite-sized, and NOT ask for too much at once. Make sure to NOT repeat any questions from the above conversation log. Answer in the following format:\nATOMIC QUESTION: the atomic question and NOTHING ELSE.\nATOMIC QUESTION: ",
+
+    "answer": "Assume that you already have enough information from the above question-answer pairs to answer the patient inquiry, use the above information to produce a factual conclusion. Respond with the correct letter choice (A, B, C, or D) and NOTHING ELSE.\nLETTER CHOICE: ",
+    
+    "non_interactive": {
+        "starter": "A patient comes into the clinic presenting with a symptom as described in the statements below:",
+        "question_prompt": "Given the information from above, your task is to choose one of four options that best answers the following question: ",
+        "response": "To the best of your ability, answer with ONLY the correct letter choice and nothing else."
+    },
+
+    "curr_template": """A patient comes into the clinic presenting with a symptom as described in the conversation log below:
+    
+PATIENT INFORMATION: {}
+CONVERSATION LOG:
+{}
+QUESTION: {}
+OPTIONS: {}
+YOUR TASK: {}"""
+    
+}
+
+patient_system = {
+    "system": "You are a truthful assistant that understands the patient's information, and you are trying to answer questions from a medical doctor about the patient. ",
+    "header": "Below is a list of factual statements about the patient:\n",
+    "prompt": 'Which of the above atomic factual statements answers the question? If no statement answers the question, simply say "The patient cannot answer this question, please do not ask this question again." Answer only what the question asks for. Do not provide any analysis, inference, or implications. Respond by selecting all statements that answer the question from above ONLY and NOTHING ELSE.',
+
+    "prompt_new": """Below is a list of factual statements about the patient:\n
+{}\n
+Which of the above atomic factual statements answers the question? If no statement answers the question, simply say "The patient cannot answer this question, please do not ask this question again." Answer only what the question asks for. Do not provide any analysis, inference, or implications. Respond with all statements that directly answer the question from above verbatim ONLY and NOTHING ELSE, with one statement on each line.
+
+Example:
+Question from the doctor: [some question]
+STATEMENTS:\n[example statement: she reports that...]\n[example statement: she has a history of...]
+
+Question from the doctor: {}
+""",
+
+    "system_first_person": "You are a patient with a list of symptoms, and you task is to truthfully answer questions from a medical doctor. ",
+    "header_first_person": "Below is a list of atomic facts about you, use ONLY the information in this list and answer the doctor's question.",
+    "prompt_first_person": """Which of the above atomic factual statements are the best answer to the question? Select at most two statements. If no statement answers the question, simply say "The patient cannot answer this question, please do not ask this question again." Do not provide any analysis, inference, or implications. Respond by reciting the matching statements, then convert the selected statements into first person perspective as if you are the patient but keep the same information. Generate your answer in this format:
+
+STATEMENTS: 
+FIRST PERSON: """
+}
+
+conformal_scores = {
+    "prompt_score": "Given the information from above, your task is to assign a likelihood score to each option. Respond with the probability as a float from 0 to 1 and NOTHING ELSE. Respond in the following format:\nA: 0.0\nB: 0.0\nC: 0.0\nD: 0.0",
+}