Training in progress, step 850, checkpoint

checkpoint-850/README.md

@@ -0,0 +1,202 @@
+---
+base_model: microsoft/Phi-3-mini-4k-instruct
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.14.0

checkpoint-850/adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "microsoft/Phi-3-mini-4k-instruct",
+  "bias": "none",
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "qkv_proj",
+    "gate_up_proj",
+    "down_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-850/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fd2fa1763366d477eb951962b2fdba14f6ba0ea026792ae67ce26ac8b2f916b
+size 25200088

checkpoint-850/global_step850/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b69aa2a0c31f9bb784767511591137482bb98bd2ec19f852bcc4305478f2a4b7
+size 18881328

checkpoint-850/global_step850/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8916d054bd36a3389788d05885a44fc84da08043a1e62bb7db1f8ef9b1165c28
+size 18881328

checkpoint-850/global_step850/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b797f524de2affa5c66530d5a8ffebb5b92ddbe02504ed958dfd93edc6ee1cbe
+size 18881328

checkpoint-850/global_step850/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a43b72a250729ef74e0080eb956f243f4464cac3c0bd4a6a69f377b2d890a576
+size 18881392

checkpoint-850/global_step850/bf16_zero_pp_rank_4_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a31c8a7cb888fdc152aac3b9e55cf947a6e9e223c60e2e649b1885609df0d790
+size 18881392

checkpoint-850/global_step850/bf16_zero_pp_rank_5_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65dfa6ce99383f2bb08a7495f1323685aa81f108bd4f54af0db9e29486849ad3
+size 18881392

checkpoint-850/global_step850/bf16_zero_pp_rank_6_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64b7b797c3b6b7be231487b80f4b234e8f3cc3cd4bc2f13d8f58f0cafcc07956
+size 18881392

checkpoint-850/global_step850/bf16_zero_pp_rank_7_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f324e9cd30ce7df7dc84d397ea3a16ff5533faebb7d2945adc03c7ea7f899910
+size 18881392

checkpoint-850/global_step850/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71ef7b64999e3c8fe2ef2bb7c33ec24ee914b6748bd90c8de0a078af8bb8baa2
+size 25379244

checkpoint-850/latest

@@ -0,0 +1 @@
+global_step850

checkpoint-850/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1754ce1fea08e0a1abf50b88b05ad2235accf247d46d7ee2f8c08c6670f73f31
+size 15984

checkpoint-850/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae26017d4550577988f9e10089ab5b71db8da5c695439c0a0fea91d6a1fd0704
+size 15984

checkpoint-850/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d1f128b23b661bf875e117cc47a5648d99e77550cfacf4588ce64a1dd7dbde3
+size 15984

checkpoint-850/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaf41da8bd40bcccaff03238fa84745187c3a9d568a9b5f691e9996625af1de6
+size 15984

checkpoint-850/rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a50ddb223b7bd2b99f1b2554cda38ae044aac0f187628b6ded5c4d407979e294
+size 15984

checkpoint-850/rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6bf0de30b7a6e43c74608e8f1fa3b7d38bb356d58e402c397bc6ad56aa95795
+size 15984

checkpoint-850/rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd016a3c3e3ba2a5ae38a6d0f24920c1961e6c3882d668aaebde5a2d6e1459fb
+size 15984

checkpoint-850/rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4e5c265f62dd45b87e17d9c102ed3afb1ecc9d2d1466b032139f4181be9bfb9
+size 15984

checkpoint-850/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e018ec51f51e576ee0ab0d945ed5b24487041ac34863e79688b43f75d30d7673
+size 1064

checkpoint-850/special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-850/tokenizer_config.json

@@ -0,0 +1,133 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": false
+    },
+    "32000": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32001": {
+      "content": "<|assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32002": {
+      "content": "<|placeholder1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32003": {
+      "content": "<|placeholder2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32004": {
+      "content": "<|placeholder3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32005": {
+      "content": "<|placeholder4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32006": {
+      "content": "<|system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32007": {
+      "content": "<|end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32008": {
+      "content": "<|placeholder5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32009": {
+      "content": "<|placeholder6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32010": {
+      "content": "<|user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "chat_template": "{% set system_message = 'You are a helpful AI assistant.' %}{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{% endif %}{% if system_message is defined %}{{ '<s>' + '<|system|>\n' + system_message + '<|end|>\n' }}{% endif %}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|user|>\n' + content + '<|end|>\n<|assistant|>\n' }}{% elif message['role'] == 'assistant' %}{{ content + '<|end|>' + '\n' }}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|end|>",
+  "extra_special_tokens": {},
+  "legacy": false,
+  "model_max_length": 4096,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "split_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

checkpoint-850/trainer_state.json

@@ -0,0 +1,1580 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.906183368869936,
+  "eval_steps": 50,
+  "global_step": 850,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.010660980810234541,
+      "grad_norm": 0.051327500492334366,
+      "learning_rate": 4.999451708687114e-06,
+      "logits/chosen": 14.755006790161133,
+      "logits/rejected": 14.735244750976562,
+      "logps/chosen": -0.29377540946006775,
+      "logps/rejected": -0.30969956517219543,
+      "loss": 0.952,
+      "rewards/accuracies": 0.4375,
+      "rewards/chosen": -0.44066309928894043,
+      "rewards/margins": 0.023886267095804214,
+      "rewards/rejected": -0.46454939246177673,
+      "step": 10
+    },
+    {
+      "epoch": 0.021321961620469083,
+      "grad_norm": 0.04346882924437523,
+      "learning_rate": 4.997807075247147e-06,
+      "logits/chosen": 14.513801574707031,
+      "logits/rejected": 14.946454048156738,
+      "logps/chosen": -0.27995699644088745,
+      "logps/rejected": -0.30138006806373596,
+      "loss": 0.9726,
+      "rewards/accuracies": 0.4124999940395355,
+      "rewards/chosen": -0.4199354648590088,
+      "rewards/margins": 0.03213457390666008,
+      "rewards/rejected": -0.45207005739212036,
+      "step": 20
+    },
+    {
+      "epoch": 0.031982942430703626,
+      "grad_norm": 0.05228634551167488,
+      "learning_rate": 4.9950668210706795e-06,
+      "logits/chosen": 14.266324043273926,
+      "logits/rejected": 14.423965454101562,
+      "logps/chosen": -0.2919609546661377,
+      "logps/rejected": -0.32358455657958984,
+      "loss": 0.9622,
+      "rewards/accuracies": 0.5,
+      "rewards/chosen": -0.43794146180152893,
+      "rewards/margins": 0.047435395419597626,
+      "rewards/rejected": -0.48537683486938477,
+      "step": 30
+    },
+    {
+      "epoch": 0.042643923240938165,
+      "grad_norm": 0.05487598106265068,
+      "learning_rate": 4.9912321481237616e-06,
+      "logits/chosen": 14.965211868286133,
+      "logits/rejected": 15.058088302612305,
+      "logps/chosen": -0.277716726064682,
+      "logps/rejected": -0.3055034577846527,
+      "loss": 0.9403,
+      "rewards/accuracies": 0.4000000059604645,
+      "rewards/chosen": -0.4165751039981842,
+      "rewards/margins": 0.04168009012937546,
+      "rewards/rejected": -0.4582551419734955,
+      "step": 40
+    },
+    {
+      "epoch": 0.053304904051172705,
+      "grad_norm": 0.057255037128925323,
+      "learning_rate": 4.986304738420684e-06,
+      "logits/chosen": 14.539288520812988,
+      "logits/rejected": 15.174041748046875,
+      "logps/chosen": -0.26362231373786926,
+      "logps/rejected": -0.3325727581977844,
+      "loss": 0.9588,
+      "rewards/accuracies": 0.4749999940395355,
+      "rewards/chosen": -0.3954334557056427,
+      "rewards/margins": 0.10342560708522797,
+      "rewards/rejected": -0.49885907769203186,
+      "step": 50
+    },
+    {
+      "epoch": 0.053304904051172705,
+      "eval_logits/chosen": 14.618952751159668,
+      "eval_logits/rejected": 15.176809310913086,
+      "eval_logps/chosen": -0.2685677409172058,
+      "eval_logps/rejected": -0.3283654451370239,
+      "eval_loss": 0.9551004767417908,
+      "eval_rewards/accuracies": 0.5131579041481018,
+      "eval_rewards/chosen": -0.4028516113758087,
+      "eval_rewards/margins": 0.08969658613204956,
+      "eval_rewards/rejected": -0.4925481975078583,
+      "eval_runtime": 21.4453,
+      "eval_samples_per_second": 28.305,
+      "eval_steps_per_second": 3.544,
+      "step": 50
+    },
+    {
+      "epoch": 0.06396588486140725,
+      "grad_norm": 0.05227242782711983,
+      "learning_rate": 4.980286753286196e-06,
+      "logits/chosen": 14.787714004516602,
+      "logits/rejected": 15.379422187805176,
+      "logps/chosen": -0.3143109679222107,
+      "logps/rejected": -0.3425135612487793,
+      "loss": 0.9636,
+      "rewards/accuracies": 0.4625000059604645,
+      "rewards/chosen": -0.4714665412902832,
+      "rewards/margins": 0.042303841561079025,
+      "rewards/rejected": -0.513770341873169,
+      "step": 60
+    },
+    {
+      "epoch": 0.07462686567164178,
+      "grad_norm": 0.0658508762717247,
+      "learning_rate": 4.973180832407471e-06,
+      "logits/chosen": 15.149365425109863,
+      "logits/rejected": 15.115835189819336,
+      "logps/chosen": -0.31501108407974243,
+      "logps/rejected": -0.2854115962982178,
+      "loss": 0.9677,
+      "rewards/accuracies": 0.36250001192092896,
+      "rewards/chosen": -0.47251659631729126,
+      "rewards/margins": -0.04439922422170639,
+      "rewards/rejected": -0.4281173646450043,
+      "step": 70
+    },
+    {
+      "epoch": 0.08528784648187633,
+      "grad_norm": 0.06567618995904922,
+      "learning_rate": 4.964990092676263e-06,
+      "logits/chosen": 15.393908500671387,
+      "logits/rejected": 15.454248428344727,
+      "logps/chosen": -0.31166282296180725,
+      "logps/rejected": -0.3178747594356537,
+      "loss": 0.9609,
+      "rewards/accuracies": 0.42500001192092896,
+      "rewards/chosen": -0.4674941897392273,
+      "rewards/margins": 0.009317949414253235,
+      "rewards/rejected": -0.47681212425231934,
+      "step": 80
+    },
+    {
+      "epoch": 0.09594882729211088,
+      "grad_norm": 0.07566913962364197,
+      "learning_rate": 4.9557181268217225e-06,
+      "logits/chosen": 15.229632377624512,
+      "logits/rejected": 15.477168083190918,
+      "logps/chosen": -0.3294064998626709,
+      "logps/rejected": -0.3528878390789032,
+      "loss": 0.9587,
+      "rewards/accuracies": 0.4749999940395355,
+      "rewards/chosen": -0.49410971999168396,
+      "rewards/margins": 0.03522203490138054,
+      "rewards/rejected": -0.5293318033218384,
+      "step": 90
+    },
+    {
+      "epoch": 0.10660980810234541,
+      "grad_norm": 0.09082464128732681,
+      "learning_rate": 4.9453690018345144e-06,
+      "logits/chosen": 14.481330871582031,
+      "logits/rejected": 15.092982292175293,
+      "logps/chosen": -0.2656436562538147,
+      "logps/rejected": -0.33982905745506287,
+      "loss": 0.9548,
+      "rewards/accuracies": 0.5625,
+      "rewards/chosen": -0.39846545457839966,
+      "rewards/margins": 0.11127817630767822,
+      "rewards/rejected": -0.5097435712814331,
+      "step": 100
+    },
+    {
+      "epoch": 0.10660980810234541,
+      "eval_logits/chosen": 14.7100830078125,
+      "eval_logits/rejected": 15.274725914001465,
+      "eval_logps/chosen": -0.26462864875793457,
+      "eval_logps/rejected": -0.331702321767807,
+      "eval_loss": 0.947841465473175,
+      "eval_rewards/accuracies": 0.5394737124443054,
+      "eval_rewards/chosen": -0.39694297313690186,
+      "eval_rewards/margins": 0.10061051696538925,
+      "eval_rewards/rejected": -0.4975534677505493,
+      "eval_runtime": 21.4421,
+      "eval_samples_per_second": 28.309,
+      "eval_steps_per_second": 3.544,
+      "step": 100
+    },
+    {
+      "epoch": 0.11727078891257996,
+      "grad_norm": 0.20198923349380493,
+      "learning_rate": 4.933947257182901e-06,
+      "logits/chosen": 14.932653427124023,
+      "logits/rejected": 15.476409912109375,
+      "logps/chosen": -0.27830976247787476,
+      "logps/rejected": -0.34150317311286926,
+      "loss": 0.9487,
+      "rewards/accuracies": 0.550000011920929,
+      "rewards/chosen": -0.41746464371681213,
+      "rewards/margins": 0.09479012340307236,
+      "rewards/rejected": -0.5122548341751099,
+      "step": 110
+    },
+    {
+      "epoch": 0.1279317697228145,
+      "grad_norm": 0.31938356161117554,
+      "learning_rate": 4.921457902821578e-06,
+      "logits/chosen": 15.280967712402344,
+      "logits/rejected": 15.5416259765625,
+      "logps/chosen": -0.2816022038459778,
+      "logps/rejected": -0.3262938857078552,
+      "loss": 0.9483,
+      "rewards/accuracies": 0.4749999940395355,
+      "rewards/chosen": -0.4224032461643219,
+      "rewards/margins": 0.06703753769397736,
+      "rewards/rejected": -0.48944082856178284,
+      "step": 120
+    },
+    {
+      "epoch": 0.13859275053304904,
+      "grad_norm": 0.12567812204360962,
+      "learning_rate": 4.907906416994146e-06,
+      "logits/chosen": 14.967382431030273,
+      "logits/rejected": 15.351877212524414,
+      "logps/chosen": -0.3148510456085205,
+      "logps/rejected": -0.3488944172859192,
+      "loss": 0.957,
+      "rewards/accuracies": 0.48750001192092896,
+      "rewards/chosen": -0.47227659821510315,
+      "rewards/margins": 0.05106503888964653,
+      "rewards/rejected": -0.5233416557312012,
+      "step": 130
+    },
+    {
+      "epoch": 0.14925373134328357,
+      "grad_norm": 0.09151162207126617,
+      "learning_rate": 4.893298743830168e-06,
+      "logits/chosen": 14.900466918945312,
+      "logits/rejected": 15.075350761413574,
+      "logps/chosen": -0.2766302227973938,
+      "logps/rejected": -0.312236487865448,
+      "loss": 0.9373,
+      "rewards/accuracies": 0.42500001192092896,
+      "rewards/chosen": -0.4149452745914459,
+      "rewards/margins": 0.05340944975614548,
+      "rewards/rejected": -0.4683547616004944,
+      "step": 140
+    },
+    {
+      "epoch": 0.15991471215351813,
+      "grad_norm": 0.1259378045797348,
+      "learning_rate": 4.8776412907378845e-06,
+      "logits/chosen": 14.528109550476074,
+      "logits/rejected": 14.861102104187012,
+      "logps/chosen": -0.2683579921722412,
+      "logps/rejected": -0.33838269114494324,
+      "loss": 0.9388,
+      "rewards/accuracies": 0.5375000238418579,
+      "rewards/chosen": -0.40253695845603943,
+      "rewards/margins": 0.10503707826137543,
+      "rewards/rejected": -0.5075740218162537,
+      "step": 150
+    },
+    {
+      "epoch": 0.15991471215351813,
+      "eval_logits/chosen": 14.12246036529541,
+      "eval_logits/rejected": 14.733266830444336,
+      "eval_logps/chosen": -0.2611957788467407,
+      "eval_logps/rejected": -0.3492279350757599,
+      "eval_loss": 0.9302574396133423,
+      "eval_rewards/accuracies": 0.5657894611358643,
+      "eval_rewards/chosen": -0.3917936384677887,
+      "eval_rewards/margins": 0.13204820454120636,
+      "eval_rewards/rejected": -0.5238418579101562,
+      "eval_runtime": 21.4406,
+      "eval_samples_per_second": 28.311,
+      "eval_steps_per_second": 3.545,
+      "step": 150
+    },
+    {
+      "epoch": 0.17057569296375266,
+      "grad_norm": 0.11400051414966583,
+      "learning_rate": 4.860940925593703e-06,
+      "logits/chosen": 14.4571533203125,
+      "logits/rejected": 14.769159317016602,
+      "logps/chosen": -0.31032469868659973,
+      "logps/rejected": -0.34650668501853943,
+      "loss": 0.9396,
+      "rewards/accuracies": 0.512499988079071,
+      "rewards/chosen": -0.4654870927333832,
+      "rewards/margins": 0.05427298694849014,
+      "rewards/rejected": -0.519760012626648,
+      "step": 160
+    },
+    {
+      "epoch": 0.1812366737739872,
+      "grad_norm": 0.1102401539683342,
+      "learning_rate": 4.84320497372973e-06,
+      "logits/chosen": 13.959765434265137,
+      "logits/rejected": 14.27458381652832,
+      "logps/chosen": -0.2744378447532654,
+      "logps/rejected": -0.35702812671661377,
+      "loss": 0.9222,
+      "rewards/accuracies": 0.5375000238418579,
+      "rewards/chosen": -0.41165676712989807,
+      "rewards/margins": 0.12388546764850616,
+      "rewards/rejected": -0.5355421900749207,
+      "step": 170
+    },
+    {
+      "epoch": 0.19189765458422176,
+      "grad_norm": 0.14721031486988068,
+      "learning_rate": 4.824441214720629e-06,
+      "logits/chosen": 13.54602336883545,
+      "logits/rejected": 14.076690673828125,
+      "logps/chosen": -0.2713850140571594,
+      "logps/rejected": -0.40618976950645447,
+      "loss": 0.9052,
+      "rewards/accuracies": 0.675000011920929,
+      "rewards/chosen": -0.40707746148109436,
+      "rewards/margins": 0.20220720767974854,
+      "rewards/rejected": -0.6092846989631653,
+      "step": 180
+    },
+    {
+      "epoch": 0.2025586353944563,
+      "grad_norm": 0.1756824553012848,
+      "learning_rate": 4.804657878971252e-06,
+      "logits/chosen": 12.6314697265625,
+      "logits/rejected": 13.246849060058594,
+      "logps/chosen": -0.27216213941574097,
+      "logps/rejected": -0.4351380467414856,
+      "loss": 0.8996,
+      "rewards/accuracies": 0.6875,
+      "rewards/chosen": -0.40824323892593384,
+      "rewards/margins": 0.24446387588977814,
+      "rewards/rejected": -0.652707040309906,
+      "step": 190
+    },
+    {
+      "epoch": 0.21321961620469082,
+      "grad_norm": 0.15476027131080627,
+      "learning_rate": 4.783863644106502e-06,
+      "logits/chosen": 12.40199089050293,
+      "logits/rejected": 12.966108322143555,
+      "logps/chosen": -0.303610622882843,
+      "logps/rejected": -0.423031747341156,
+      "loss": 0.9015,
+      "rewards/accuracies": 0.5375000238418579,
+      "rewards/chosen": -0.4554159641265869,
+      "rewards/margins": 0.17913168668746948,
+      "rewards/rejected": -0.6345476508140564,
+      "step": 200
+    },
+    {
+      "epoch": 0.21321961620469082,
+      "eval_logits/chosen": 11.887229919433594,
+      "eval_logits/rejected": 12.5900239944458,
+      "eval_logps/chosen": -0.269090861082077,
+      "eval_logps/rejected": -0.42408913373947144,
+      "eval_loss": 0.8796805143356323,
+      "eval_rewards/accuracies": 0.6447368264198303,
+      "eval_rewards/chosen": -0.40363630652427673,
+      "eval_rewards/margins": 0.23249731957912445,
+      "eval_rewards/rejected": -0.6361336708068848,
+      "eval_runtime": 21.4455,
+      "eval_samples_per_second": 28.304,
+      "eval_steps_per_second": 3.544,
+      "step": 200
+    },
+    {
+      "epoch": 0.22388059701492538,
+      "grad_norm": 0.18212148547172546,
+      "learning_rate": 4.762067631165049e-06,
+      "logits/chosen": 12.375594139099121,
+      "logits/rejected": 12.701678276062012,
+      "logps/chosen": -0.3136894702911377,
+      "logps/rejected": -0.3944609761238098,
+      "loss": 0.8898,
+      "rewards/accuracies": 0.4625000059604645,
+      "rewards/chosen": -0.47053417563438416,
+      "rewards/margins": 0.12115727365016937,
+      "rewards/rejected": -0.5916914939880371,
+      "step": 210
+    },
+    {
+      "epoch": 0.2345415778251599,
+      "grad_norm": 0.5440058708190918,
+      "learning_rate": 4.7392794005985324e-06,
+      "logits/chosen": 11.23914909362793,
+      "logits/rejected": 11.926396369934082,
+      "logps/chosen": -0.3077571392059326,
+      "logps/rejected": -0.43772149085998535,
+      "loss": 0.8806,
+      "rewards/accuracies": 0.5375000238418579,
+      "rewards/chosen": -0.4616357684135437,
+      "rewards/margins": 0.19494646787643433,
+      "rewards/rejected": -0.656582236289978,
+      "step": 220
+    },
+    {
+      "epoch": 0.24520255863539445,
+      "grad_norm": 0.5628307461738586,
+      "learning_rate": 4.715508948078037e-06,
+      "logits/chosen": 11.177714347839355,
+      "logits/rejected": 11.534266471862793,
+      "logps/chosen": -0.31991320848464966,
+      "logps/rejected": -0.4394511282444,
+      "loss": 0.8778,
+      "rewards/accuracies": 0.550000011920929,
+      "rewards/chosen": -0.4798697829246521,
+      "rewards/margins": 0.17930689454078674,
+      "rewards/rejected": -0.6591767072677612,
+      "step": 230
+    },
+    {
+      "epoch": 0.255863539445629,
+      "grad_norm": 0.40485626459121704,
+      "learning_rate": 4.690766700109659e-06,
+      "logits/chosen": 10.132668495178223,
+      "logits/rejected": 10.29063606262207,
+      "logps/chosen": -0.3195653557777405,
+      "logps/rejected": -0.47949132323265076,
+      "loss": 0.8551,
+      "rewards/accuracies": 0.5874999761581421,
+      "rewards/chosen": -0.47934800386428833,
+      "rewards/margins": 0.23988890647888184,
+      "rewards/rejected": -0.7192369699478149,
+      "step": 240
+    },
+    {
+      "epoch": 0.26652452025586354,
+      "grad_norm": 0.6199322938919067,
+      "learning_rate": 4.665063509461098e-06,
+      "logits/chosen": 8.781888008117676,
+      "logits/rejected": 9.237382888793945,
+      "logps/chosen": -0.3370448052883148,
+      "logps/rejected": -0.610824465751648,
+      "loss": 0.8416,
+      "rewards/accuracies": 0.7250000238418579,
+      "rewards/chosen": -0.505567193031311,
+      "rewards/margins": 0.4106695055961609,
+      "rewards/rejected": -0.9162367582321167,
+      "step": 250
+    },
+    {
+      "epoch": 0.26652452025586354,
+      "eval_logits/chosen": 8.437722206115723,
+      "eval_logits/rejected": 8.843962669372559,
+      "eval_logps/chosen": -0.3058585226535797,
+      "eval_logps/rejected": -0.582990825176239,
+      "eval_loss": 0.8036603331565857,
+      "eval_rewards/accuracies": 0.6447368264198303,
+      "eval_rewards/chosen": -0.4587877094745636,
+      "eval_rewards/margins": 0.4156985878944397,
+      "eval_rewards/rejected": -0.8744862079620361,
+      "eval_runtime": 21.4423,
+      "eval_samples_per_second": 28.308,
+      "eval_steps_per_second": 3.544,
+      "step": 250
+    },
+    {
+      "epoch": 0.2771855010660981,
+      "grad_norm": 0.3213505744934082,
+      "learning_rate": 4.638410650401267e-06,
+      "logits/chosen": 7.914826393127441,
+      "logits/rejected": 8.010818481445312,
+      "logps/chosen": -0.3556877374649048,
+      "logps/rejected": -0.7540119886398315,
+      "loss": 0.7811,
+      "rewards/accuracies": 0.6499999761581421,
+      "rewards/chosen": -0.5335315465927124,
+      "rewards/margins": 0.5974863171577454,
+      "rewards/rejected": -1.1310179233551025,
+      "step": 260
+    },
+    {
+      "epoch": 0.2878464818763326,
+      "grad_norm": 1.0119378566741943,
+      "learning_rate": 4.610819813755038e-06,
+      "logits/chosen": 7.584845542907715,
+      "logits/rejected": 7.812608242034912,
+      "logps/chosen": -0.3649575412273407,
+      "logps/rejected": -0.8042632937431335,
+      "loss": 0.7391,
+      "rewards/accuracies": 0.5249999761581421,
+      "rewards/chosen": -0.5474363565444946,
+      "rewards/margins": 0.6589586734771729,
+      "rewards/rejected": -1.206395149230957,
+      "step": 270
+    },
+    {
+      "epoch": 0.29850746268656714,
+      "grad_norm": 0.5339816808700562,
+      "learning_rate": 4.582303101775249e-06,
+      "logits/chosen": 6.687758445739746,
+      "logits/rejected": 6.233181476593018,
+      "logps/chosen": -0.415935218334198,
+      "logps/rejected": -1.2987438440322876,
+      "loss": 0.7419,
+      "rewards/accuracies": 0.6000000238418579,
+      "rewards/chosen": -0.6239027976989746,
+      "rewards/margins": 1.3242127895355225,
+      "rewards/rejected": -1.9481157064437866,
+      "step": 280
+    },
+    {
+      "epoch": 0.3091684434968017,
+      "grad_norm": 0.3514000475406647,
+      "learning_rate": 4.55287302283426e-06,
+      "logits/chosen": 6.2503981590271,
+      "logits/rejected": 5.798542499542236,
+      "logps/chosen": -0.4319223463535309,
+      "logps/rejected": -1.2257453203201294,
+      "loss": 0.7235,
+      "rewards/accuracies": 0.6000000238418579,
+      "rewards/chosen": -0.6478835344314575,
+      "rewards/margins": 1.1907342672348022,
+      "rewards/rejected": -1.8386180400848389,
+      "step": 290
+    },
+    {
+      "epoch": 0.31982942430703626,
+      "grad_norm": 0.6761008501052856,
+      "learning_rate": 4.522542485937369e-06,
+      "logits/chosen": 4.4480695724487305,
+      "logits/rejected": 4.290585994720459,
+      "logps/chosen": -0.42002564668655396,
+      "logps/rejected": -1.4215493202209473,
+      "loss": 0.7058,
+      "rewards/accuracies": 0.637499988079071,
+      "rewards/chosen": -0.6300384402275085,
+      "rewards/margins": 1.5022855997085571,
+      "rewards/rejected": -2.132323980331421,
+      "step": 300
+    },
+    {
+      "epoch": 0.31982942430703626,
+      "eval_logits/chosen": 4.789332389831543,
+      "eval_logits/rejected": 4.481485366821289,
+      "eval_logps/chosen": -0.4049508571624756,
+      "eval_logps/rejected": -1.395646095275879,
+      "eval_loss": 0.6695442199707031,
+      "eval_rewards/accuracies": 0.6710526347160339,
+      "eval_rewards/chosen": -0.6074262857437134,
+      "eval_rewards/margins": 1.4860429763793945,
+      "eval_rewards/rejected": -2.0934693813323975,
+      "eval_runtime": 21.4397,
+      "eval_samples_per_second": 28.312,
+      "eval_steps_per_second": 3.545,
+      "step": 300
+    },
+    {
+      "epoch": 0.3304904051172708,
+      "grad_norm": 0.44682690501213074,
+      "learning_rate": 4.491324795060491e-06,
+      "logits/chosen": 5.487166404724121,
+      "logits/rejected": 4.501384258270264,
+      "logps/chosen": -0.5215579867362976,
+      "logps/rejected": -1.7223398685455322,
+      "loss": 0.6988,
+      "rewards/accuracies": 0.612500011920929,
+      "rewards/chosen": -0.7823370695114136,
+      "rewards/margins": 1.8011726140975952,
+      "rewards/rejected": -2.5835094451904297,
+      "step": 310
+    },
+    {
+      "epoch": 0.3411513859275053,
+      "grad_norm": 0.41085830330848694,
+      "learning_rate": 4.4592336433146e-06,
+      "logits/chosen": 4.162590026855469,
+      "logits/rejected": 2.876271963119507,
+      "logps/chosen": -0.5402930974960327,
+      "logps/rejected": -1.7925996780395508,
+      "loss": 0.6811,
+      "rewards/accuracies": 0.625,
+      "rewards/chosen": -0.8104397058486938,
+      "rewards/margins": 1.8784599304199219,
+      "rewards/rejected": -2.688899517059326,
+      "step": 320
+    },
+    {
+      "epoch": 0.35181236673773986,
+      "grad_norm": 0.5611584186553955,
+      "learning_rate": 4.426283106939474e-06,
+      "logits/chosen": 4.088540077209473,
+      "logits/rejected": 3.081679582595825,
+      "logps/chosen": -0.541223406791687,
+      "logps/rejected": -1.9464069604873657,
+      "loss": 0.6614,
+      "rewards/accuracies": 0.699999988079071,
+      "rewards/chosen": -0.8118351101875305,
+      "rewards/margins": 2.1077752113342285,
+      "rewards/rejected": -2.919610023498535,
+      "step": 330
+    },
+    {
+      "epoch": 0.3624733475479744,
+      "grad_norm": 4.05828857421875,
+      "learning_rate": 4.3924876391293915e-06,
+      "logits/chosen": 3.3937134742736816,
+      "logits/rejected": 2.4182538986206055,
+      "logps/chosen": -0.6656067967414856,
+      "logps/rejected": -1.5255868434906006,
+      "loss": 0.6583,
+      "rewards/accuracies": 0.574999988079071,
+      "rewards/chosen": -0.9984102249145508,
+      "rewards/margins": 1.2899701595306396,
+      "rewards/rejected": -2.2883803844451904,
+      "step": 340
+    },
+    {
+      "epoch": 0.373134328358209,
+      "grad_norm": 0.8311880230903625,
+      "learning_rate": 4.357862063693486e-06,
+      "logits/chosen": 2.503194570541382,
+      "logits/rejected": 1.5284960269927979,
+      "logps/chosen": -0.6593035459518433,
+      "logps/rejected": -2.211193323135376,
+      "loss": 0.5911,
+      "rewards/accuracies": 0.699999988079071,
+      "rewards/chosen": -0.9889553189277649,
+      "rewards/margins": 2.3278346061706543,
+      "rewards/rejected": -3.3167896270751953,
+      "step": 350
+    },
+    {
+      "epoch": 0.373134328358209,
+      "eval_logits/chosen": 2.556962728500366,
+      "eval_logits/rejected": 1.830418586730957,
+      "eval_logps/chosen": -0.6546408534049988,
+      "eval_logps/rejected": -1.9014692306518555,
+      "eval_loss": 0.5961893200874329,
+      "eval_rewards/accuracies": 0.6842105388641357,
+      "eval_rewards/chosen": -0.9819613099098206,
+      "eval_rewards/margins": 1.8702424764633179,
+      "eval_rewards/rejected": -2.852203845977783,
+      "eval_runtime": 21.4393,
+      "eval_samples_per_second": 28.312,
+      "eval_steps_per_second": 3.545,
+      "step": 350
+    },
+    {
+      "epoch": 0.3837953091684435,
+      "grad_norm": 1.4237236976623535,
+      "learning_rate": 4.322421568553529e-06,
+      "logits/chosen": 3.0001542568206787,
+      "logits/rejected": 1.9715242385864258,
+      "logps/chosen": -0.8050466775894165,
+      "logps/rejected": -2.2938907146453857,
+      "loss": 0.58,
+      "rewards/accuracies": 0.737500011920929,
+      "rewards/chosen": -1.20756995677948,
+      "rewards/margins": 2.2332661151885986,
+      "rewards/rejected": -3.440835952758789,
+      "step": 360
+    },
+    {
+      "epoch": 0.39445628997867804,
+      "grad_norm": 2.2651443481445312,
+      "learning_rate": 4.286181699082008e-06,
+      "logits/chosen": 2.7526040077209473,
+      "logits/rejected": 2.05066180229187,
+      "logps/chosen": -1.6301355361938477,
+      "logps/rejected": -2.9630703926086426,
+      "loss": 0.5823,
+      "rewards/accuracies": 0.75,
+      "rewards/chosen": -2.4452033042907715,
+      "rewards/margins": 1.999402642250061,
+      "rewards/rejected": -4.444605827331543,
+      "step": 370
+    },
+    {
+      "epoch": 0.4051172707889126,
+      "grad_norm": 1.9120367765426636,
+      "learning_rate": 4.249158351283414e-06,
+      "logits/chosen": 1.9757938385009766,
+      "logits/rejected": 1.5915673971176147,
+      "logps/chosen": -2.063323497772217,
+      "logps/rejected": -2.899749755859375,
+      "loss": 0.5675,
+      "rewards/accuracies": 0.737500011920929,
+      "rewards/chosen": -3.094984769821167,
+      "rewards/margins": 1.2546398639678955,
+      "rewards/rejected": -4.3496246337890625,
+      "step": 380
+    },
+    {
+      "epoch": 0.4157782515991471,
+      "grad_norm": 3.0018720626831055,
+      "learning_rate": 4.211367764821722e-06,
+      "logits/chosen": 2.541440486907959,
+      "logits/rejected": 1.7436832189559937,
+      "logps/chosen": -2.279510736465454,
+      "logps/rejected": -3.3447775840759277,
+      "loss": 0.4969,
+      "rewards/accuracies": 0.8500000238418579,
+      "rewards/chosen": -3.4192657470703125,
+      "rewards/margins": 1.5979007482528687,
+      "rewards/rejected": -5.017167091369629,
+      "step": 390
+    },
+    {
+      "epoch": 0.42643923240938164,
+      "grad_norm": 1.9656275510787964,
+      "learning_rate": 4.172826515897146e-06,
+      "logits/chosen": 1.6748476028442383,
+      "logits/rejected": 1.0921740531921387,
+      "logps/chosen": -2.147991180419922,
+      "logps/rejected": -3.380042314529419,
+      "loss": 0.5135,
+      "rewards/accuracies": 0.8374999761581421,
+      "rewards/chosen": -3.221986770629883,
+      "rewards/margins": 1.8480768203735352,
+      "rewards/rejected": -5.07006311416626,
+      "step": 400
+    },
+    {
+      "epoch": 0.42643923240938164,
+      "eval_logits/chosen": 2.210231065750122,
+      "eval_logits/rejected": 1.679926872253418,
+      "eval_logps/chosen": -2.044506788253784,
+      "eval_logps/rejected": -3.713956356048584,
+      "eval_loss": 0.47455134987831116,
+      "eval_rewards/accuracies": 0.9342105388641357,
+      "eval_rewards/chosen": -3.0667598247528076,
+      "eval_rewards/margins": 2.5041754245758057,
+      "eval_rewards/rejected": -5.570935249328613,
+      "eval_runtime": 21.4401,
+      "eval_samples_per_second": 28.311,
+      "eval_steps_per_second": 3.545,
+      "step": 400
+    },
+    {
+      "epoch": 0.43710021321961623,
+      "grad_norm": 2.501361131668091,
+      "learning_rate": 4.133551509975264e-06,
+      "logits/chosen": 1.9820306301116943,
+      "logits/rejected": 1.3992068767547607,
+      "logps/chosen": -2.300197124481201,
+      "logps/rejected": -3.813164472579956,
+      "loss": 0.498,
+      "rewards/accuracies": 0.8500000238418579,
+      "rewards/chosen": -3.4502956867218018,
+      "rewards/margins": 2.2694506645202637,
+      "rewards/rejected": -5.7197465896606445,
+      "step": 410
+    },
+    {
+      "epoch": 0.44776119402985076,
+      "grad_norm": 3.828648090362549,
+      "learning_rate": 4.093559974371725e-06,
+      "logits/chosen": 2.7997095584869385,
+      "logits/rejected": 2.4387598037719727,
+      "logps/chosen": -2.687736749649048,
+      "logps/rejected": -4.425741195678711,
+      "loss": 0.4494,
+      "rewards/accuracies": 0.862500011920929,
+      "rewards/chosen": -4.031605243682861,
+      "rewards/margins": 2.607006788253784,
+      "rewards/rejected": -6.638613224029541,
+      "step": 420
+    },
+    {
+      "epoch": 0.4584221748400853,
+      "grad_norm": 2.635803461074829,
+      "learning_rate": 4.052869450695776e-06,
+      "logits/chosen": 2.942661762237549,
+      "logits/rejected": 2.019963026046753,
+      "logps/chosen": -2.98117733001709,
+      "logps/rejected": -4.717232704162598,
+      "loss": 0.4796,
+      "rewards/accuracies": 0.8500000238418579,
+      "rewards/chosen": -4.471765518188477,
+      "rewards/margins": 2.60408353805542,
+      "rewards/rejected": -7.075850009918213,
+      "step": 430
+    },
+    {
+      "epoch": 0.4690831556503198,
+      "grad_norm": 3.140829086303711,
+      "learning_rate": 4.011497787155938e-06,
+      "logits/chosen": 3.2747459411621094,
+      "logits/rejected": 2.2958083152770996,
+      "logps/chosen": -3.129321575164795,
+      "logps/rejected": -4.921725273132324,
+      "loss": 0.4468,
+      "rewards/accuracies": 0.862500011920929,
+      "rewards/chosen": -4.69398307800293,
+      "rewards/margins": 2.688605785369873,
+      "rewards/rejected": -7.3825883865356445,
+      "step": 440
+    },
+    {
+      "epoch": 0.47974413646055436,
+      "grad_norm": 2.7932240962982178,
+      "learning_rate": 3.969463130731183e-06,
+      "logits/chosen": 2.205420970916748,
+      "logits/rejected": 1.4024155139923096,
+      "logps/chosen": -2.7564563751220703,
+      "logps/rejected": -4.563851356506348,
+      "loss": 0.4073,
+      "rewards/accuracies": 0.887499988079071,
+      "rewards/chosen": -4.1346845626831055,
+      "rewards/margins": 2.711092472076416,
+      "rewards/rejected": -6.8457770347595215,
+      "step": 450
+    },
+    {
+      "epoch": 0.47974413646055436,
+      "eval_logits/chosen": 2.0136826038360596,
+      "eval_logits/rejected": 1.561701774597168,
+      "eval_logps/chosen": -2.7486908435821533,
+      "eval_logps/rejected": -4.690793514251709,
+      "eval_loss": 0.41499289870262146,
+      "eval_rewards/accuracies": 0.9210526347160339,
+      "eval_rewards/chosen": -4.123035907745361,
+      "eval_rewards/margins": 2.913153648376465,
+      "eval_rewards/rejected": -7.036189079284668,
+      "eval_runtime": 21.4387,
+      "eval_samples_per_second": 28.313,
+      "eval_steps_per_second": 3.545,
+      "step": 450
+    },
+    {
+      "epoch": 0.4904051172707889,
+      "grad_norm": 2.7059199810028076,
+      "learning_rate": 3.92678391921108e-06,
+      "logits/chosen": 2.257246494293213,
+      "logits/rejected": 1.6654322147369385,
+      "logps/chosen": -3.389554500579834,
+      "logps/rejected": -5.6951165199279785,
+      "loss": 0.4004,
+      "rewards/accuracies": 0.949999988079071,
+      "rewards/chosen": -5.084332466125488,
+      "rewards/margins": 3.4583427906036377,
+      "rewards/rejected": -8.542675018310547,
+      "step": 460
+    },
+    {
+      "epoch": 0.5010660980810234,
+      "grad_norm": 2.245579719543457,
+      "learning_rate": 3.88347887310836e-06,
+      "logits/chosen": 2.3386971950531006,
+      "logits/rejected": 2.086036205291748,
+      "logps/chosen": -3.2753937244415283,
+      "logps/rejected": -5.4362359046936035,
+      "loss": 0.3976,
+      "rewards/accuracies": 0.875,
+      "rewards/chosen": -4.913090705871582,
+      "rewards/margins": 3.241262912750244,
+      "rewards/rejected": -8.154353141784668,
+      "step": 470
+    },
+    {
+      "epoch": 0.511727078891258,
+      "grad_norm": 2.8131167888641357,
+      "learning_rate": 3.839566987447492e-06,
+      "logits/chosen": 1.6951004266738892,
+      "logits/rejected": 1.3795586824417114,
+      "logps/chosen": -3.2933483123779297,
+      "logps/rejected": -5.050060749053955,
+      "loss": 0.3982,
+      "rewards/accuracies": 0.8999999761581421,
+      "rewards/chosen": -4.940022945404053,
+      "rewards/margins": 2.635068416595459,
+      "rewards/rejected": -7.5750908851623535,
+      "step": 480
+    },
+    {
+      "epoch": 0.5223880597014925,
+      "grad_norm": 2.6465814113616943,
+      "learning_rate": 3.795067523432826e-06,
+      "logits/chosen": 3.136894702911377,
+      "logits/rejected": 2.6332411766052246,
+      "logps/chosen": -4.419378757476807,
+      "logps/rejected": -6.467301845550537,
+      "loss": 0.3974,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -6.629067420959473,
+      "rewards/margins": 3.071885585784912,
+      "rewards/rejected": -9.700952529907227,
+      "step": 490
+    },
+    {
+      "epoch": 0.5330490405117271,
+      "grad_norm": 3.6718053817749023,
+      "learning_rate": 3.7500000000000005e-06,
+      "logits/chosen": 1.681780457496643,
+      "logits/rejected": 1.0038775205612183,
+      "logps/chosen": -3.266970157623291,
+      "logps/rejected": -5.594450950622559,
+      "loss": 0.367,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -4.900455474853516,
+      "rewards/margins": 3.4912209510803223,
+      "rewards/rejected": -8.391676902770996,
+      "step": 500
+    },
+    {
+      "epoch": 0.5330490405117271,
+      "eval_logits/chosen": 2.110192060470581,
+      "eval_logits/rejected": 1.7233155965805054,
+      "eval_logps/chosen": -3.0329930782318115,
+      "eval_logps/rejected": -5.3280930519104,
+      "eval_loss": 0.387028306722641,
+      "eval_rewards/accuracies": 0.9210526347160339,
+      "eval_rewards/chosen": -4.549489498138428,
+      "eval_rewards/margins": 3.4426498413085938,
+      "eval_rewards/rejected": -7.9921393394470215,
+      "eval_runtime": 21.4417,
+      "eval_samples_per_second": 28.309,
+      "eval_steps_per_second": 3.545,
+      "step": 500
+    },
+    {
+      "epoch": 0.5437100213219617,
+      "grad_norm": 3.392418622970581,
+      "learning_rate": 3.7043841852542884e-06,
+      "logits/chosen": 3.0106852054595947,
+      "logits/rejected": 2.309483528137207,
+      "logps/chosen": -3.4964828491210938,
+      "logps/rejected": -5.438345909118652,
+      "loss": 0.3776,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -5.244723796844482,
+      "rewards/margins": 2.9127936363220215,
+      "rewards/rejected": -8.15751838684082,
+      "step": 510
+    },
+    {
+      "epoch": 0.5543710021321961,
+      "grad_norm": 3.551785707473755,
+      "learning_rate": 3.658240087799655e-06,
+      "logits/chosen": 2.8624072074890137,
+      "logits/rejected": 2.425307512283325,
+      "logps/chosen": -3.780029773712158,
+      "logps/rejected": -5.929925918579102,
+      "loss": 0.3999,
+      "rewards/accuracies": 0.887499988079071,
+      "rewards/chosen": -5.670044898986816,
+      "rewards/margins": 3.224843978881836,
+      "rewards/rejected": -8.894887924194336,
+      "step": 520
+    },
+    {
+      "epoch": 0.5650319829424307,
+      "grad_norm": 3.0529866218566895,
+      "learning_rate": 3.611587947962319e-06,
+      "logits/chosen": 2.3145668506622314,
+      "logits/rejected": 1.5088326930999756,
+      "logps/chosen": -3.1259171962738037,
+      "logps/rejected": -5.282050132751465,
+      "loss": 0.3346,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -4.688876152038574,
+      "rewards/margins": 3.234198808670044,
+      "rewards/rejected": -7.923074245452881,
+      "step": 530
+    },
+    {
+      "epoch": 0.5756929637526652,
+      "grad_norm": 1.7146470546722412,
+      "learning_rate": 3.564448228912682e-06,
+      "logits/chosen": 2.5195610523223877,
+      "logits/rejected": 2.286477565765381,
+      "logps/chosen": -4.000255107879639,
+      "logps/rejected": -6.475634574890137,
+      "loss": 0.3675,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -6.000383377075195,
+      "rewards/margins": 3.713069438934326,
+      "rewards/rejected": -9.713452339172363,
+      "step": 540
+    },
+    {
+      "epoch": 0.5863539445628998,
+      "grad_norm": 8.31704330444336,
+      "learning_rate": 3.516841607689501e-06,
+      "logits/chosen": 1.8846759796142578,
+      "logits/rejected": 1.5663390159606934,
+      "logps/chosen": -3.9287109375,
+      "logps/rejected": -6.041600704193115,
+      "loss": 0.3921,
+      "rewards/accuracies": 0.862500011920929,
+      "rewards/chosen": -5.89306640625,
+      "rewards/margins": 3.1693339347839355,
+      "rewards/rejected": -9.062400817871094,
+      "step": 550
+    },
+    {
+      "epoch": 0.5863539445628998,
+      "eval_logits/chosen": 2.0545763969421387,
+      "eval_logits/rejected": 1.7521167993545532,
+      "eval_logps/chosen": -3.2991809844970703,
+      "eval_logps/rejected": -5.87559175491333,
+      "eval_loss": 0.35785165429115295,
+      "eval_rewards/accuracies": 0.9342105388641357,
+      "eval_rewards/chosen": -4.948771953582764,
+      "eval_rewards/margins": 3.8646163940429688,
+      "eval_rewards/rejected": -8.813387870788574,
+      "eval_runtime": 21.439,
+      "eval_samples_per_second": 28.313,
+      "eval_steps_per_second": 3.545,
+      "step": 550
+    },
+    {
+      "epoch": 0.5970149253731343,
+      "grad_norm": 4.491275787353516,
+      "learning_rate": 3.4687889661302577e-06,
+      "logits/chosen": 2.9886631965637207,
+      "logits/rejected": 2.2341508865356445,
+      "logps/chosen": -4.050547122955322,
+      "logps/rejected": -6.301305294036865,
+      "loss": 0.3701,
+      "rewards/accuracies": 0.887499988079071,
+      "rewards/chosen": -6.0758209228515625,
+      "rewards/margins": 3.3761372566223145,
+      "rewards/rejected": -9.451958656311035,
+      "step": 560
+    },
+    {
+      "epoch": 0.6076759061833689,
+      "grad_norm": 2.40407395362854,
+      "learning_rate": 3.4203113817116955e-06,
+      "logits/chosen": 3.0933375358581543,
+      "logits/rejected": 2.4652345180511475,
+      "logps/chosen": -3.9168992042541504,
+      "logps/rejected": -5.933487892150879,
+      "loss": 0.3481,
+      "rewards/accuracies": 0.8999999761581421,
+      "rewards/chosen": -5.875349521636963,
+      "rewards/margins": 3.024883508682251,
+      "rewards/rejected": -8.900232315063477,
+      "step": 570
+    },
+    {
+      "epoch": 0.6183368869936035,
+      "grad_norm": 4.073615550994873,
+      "learning_rate": 3.3714301183045382e-06,
+      "logits/chosen": 2.5948500633239746,
+      "logits/rejected": 2.202971935272217,
+      "logps/chosen": -3.6740527153015137,
+      "logps/rejected": -6.205447196960449,
+      "loss": 0.3571,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -5.511078834533691,
+      "rewards/margins": 3.7970924377441406,
+      "rewards/rejected": -9.308171272277832,
+      "step": 580
+    },
+    {
+      "epoch": 0.6289978678038379,
+      "grad_norm": 2.427555799484253,
+      "learning_rate": 3.3221666168464584e-06,
+      "logits/chosen": 1.6076780557632446,
+      "logits/rejected": 1.0556285381317139,
+      "logps/chosen": -3.609313488006592,
+      "logps/rejected": -5.912892818450928,
+      "loss": 0.3666,
+      "rewards/accuracies": 0.875,
+      "rewards/chosen": -5.413969993591309,
+      "rewards/margins": 3.455368757247925,
+      "rewards/rejected": -8.869338989257812,
+      "step": 590
+    },
+    {
+      "epoch": 0.6396588486140725,
+      "grad_norm": 3.910998821258545,
+      "learning_rate": 3.272542485937369e-06,
+      "logits/chosen": 2.5494394302368164,
+      "logits/rejected": 1.9157211780548096,
+      "logps/chosen": -4.038924217224121,
+      "logps/rejected": -6.580315589904785,
+      "loss": 0.3017,
+      "rewards/accuracies": 0.862500011920929,
+      "rewards/chosen": -6.058385372161865,
+      "rewards/margins": 3.812087297439575,
+      "rewards/rejected": -9.87047290802002,
+      "step": 600
+    },
+    {
+      "epoch": 0.6396588486140725,
+      "eval_logits/chosen": 2.179999589920044,
+      "eval_logits/rejected": 1.918432593345642,
+      "eval_logps/chosen": -3.3114354610443115,
+      "eval_logps/rejected": -6.157403469085693,
+      "eval_loss": 0.3409230411052704,
+      "eval_rewards/accuracies": 0.9210526347160339,
+      "eval_rewards/chosen": -4.967153549194336,
+      "eval_rewards/margins": 4.268952369689941,
+      "eval_rewards/rejected": -9.236105918884277,
+      "eval_runtime": 21.4405,
+      "eval_samples_per_second": 28.311,
+      "eval_steps_per_second": 3.545,
+      "step": 600
+    },
+    {
+      "epoch": 0.650319829424307,
+      "grad_norm": 3.6193580627441406,
+      "learning_rate": 3.222579492361179e-06,
+      "logits/chosen": 2.680032968521118,
+      "logits/rejected": 1.9571218490600586,
+      "logps/chosen": -3.6821212768554688,
+      "logps/rejected": -6.76863956451416,
+      "loss": 0.3429,
+      "rewards/accuracies": 0.9375,
+      "rewards/chosen": -5.523181915283203,
+      "rewards/margins": 4.629776954650879,
+      "rewards/rejected": -10.152959823608398,
+      "step": 610
+    },
+    {
+      "epoch": 0.6609808102345416,
+      "grad_norm": 4.7948808670043945,
+      "learning_rate": 3.1722995515381644e-06,
+      "logits/chosen": 2.570702075958252,
+      "logits/rejected": 2.2683706283569336,
+      "logps/chosen": -4.124785423278809,
+      "logps/rejected": -6.595047950744629,
+      "loss": 0.3355,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -6.187178134918213,
+      "rewards/margins": 3.7053933143615723,
+      "rewards/rejected": -9.892572402954102,
+      "step": 620
+    },
+    {
+      "epoch": 0.6716417910447762,
+      "grad_norm": 2.727231025695801,
+      "learning_rate": 3.121724717912138e-06,
+      "logits/chosen": 2.0630054473876953,
+      "logits/rejected": 1.4909979104995728,
+      "logps/chosen": -3.6443302631378174,
+      "logps/rejected": -6.239049911499023,
+      "loss": 0.3218,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -5.466495513916016,
+      "rewards/margins": 3.8920791149139404,
+      "rewards/rejected": -9.358574867248535,
+      "step": 630
+    },
+    {
+      "epoch": 0.6823027718550106,
+      "grad_norm": 4.138846397399902,
+      "learning_rate": 3.0708771752766397e-06,
+      "logits/chosen": 1.6833502054214478,
+      "logits/rejected": 1.2245066165924072,
+      "logps/chosen": -4.318512916564941,
+      "logps/rejected": -7.222353935241699,
+      "loss": 0.2868,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -6.477769374847412,
+      "rewards/margins": 4.35576057434082,
+      "rewards/rejected": -10.833529472351074,
+      "step": 640
+    },
+    {
+      "epoch": 0.6929637526652452,
+      "grad_norm": 7.172155380249023,
+      "learning_rate": 3.019779227044398e-06,
+      "logits/chosen": 2.5162904262542725,
+      "logits/rejected": 2.0894787311553955,
+      "logps/chosen": -4.13573694229126,
+      "logps/rejected": -7.199030876159668,
+      "loss": 0.3491,
+      "rewards/accuracies": 0.9375,
+      "rewards/chosen": -6.2036051750183105,
+      "rewards/margins": 4.594940185546875,
+      "rewards/rejected": -10.798544883728027,
+      "step": 650
+    },
+    {
+      "epoch": 0.6929637526652452,
+      "eval_logits/chosen": 2.003471612930298,
+      "eval_logits/rejected": 1.8008451461791992,
+      "eval_logps/chosen": -3.2873525619506836,
+      "eval_logps/rejected": -6.338763236999512,
+      "eval_loss": 0.31618499755859375,
+      "eval_rewards/accuracies": 0.9342105388641357,
+      "eval_rewards/chosen": -4.931028842926025,
+      "eval_rewards/margins": 4.5771164894104,
+      "eval_rewards/rejected": -9.508145332336426,
+      "eval_runtime": 21.415,
+      "eval_samples_per_second": 28.345,
+      "eval_steps_per_second": 3.549,
+      "step": 650
+    },
+    {
+      "epoch": 0.7036247334754797,
+      "grad_norm": 3.3068535327911377,
+      "learning_rate": 2.9684532864643123e-06,
+      "logits/chosen": 2.852752208709717,
+      "logits/rejected": 2.3139753341674805,
+      "logps/chosen": -3.9495348930358887,
+      "logps/rejected": -6.509753227233887,
+      "loss": 0.2684,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -5.924302101135254,
+      "rewards/margins": 3.8403282165527344,
+      "rewards/rejected": -9.764630317687988,
+      "step": 660
+    },
+    {
+      "epoch": 0.7142857142857143,
+      "grad_norm": 2.737560987472534,
+      "learning_rate": 2.9169218667902562e-06,
+      "logits/chosen": 3.034759998321533,
+      "logits/rejected": 2.2936713695526123,
+      "logps/chosen": -3.855508327484131,
+      "logps/rejected": -6.571882724761963,
+      "loss": 0.2887,
+      "rewards/accuracies": 0.9375,
+      "rewards/chosen": -5.783262252807617,
+      "rewards/margins": 4.0745625495910645,
+      "rewards/rejected": -9.857824325561523,
+      "step": 670
+    },
+    {
+      "epoch": 0.7249466950959488,
+      "grad_norm": 4.387801647186279,
+      "learning_rate": 2.8652075714060296e-06,
+      "logits/chosen": 1.5638288259506226,
+      "logits/rejected": 1.1110167503356934,
+      "logps/chosen": -3.7569847106933594,
+      "logps/rejected": -6.927552223205566,
+      "loss": 0.3319,
+      "rewards/accuracies": 0.9624999761581421,
+      "rewards/chosen": -5.635476589202881,
+      "rewards/margins": 4.755851745605469,
+      "rewards/rejected": -10.391328811645508,
+      "step": 680
+    },
+    {
+      "epoch": 0.7356076759061834,
+      "grad_norm": 2.357923746109009,
+      "learning_rate": 2.813333083910761e-06,
+      "logits/chosen": 2.058987855911255,
+      "logits/rejected": 1.6591012477874756,
+      "logps/chosen": -4.124215126037598,
+      "logps/rejected": -7.349566459655762,
+      "loss": 0.2362,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -6.186322212219238,
+      "rewards/margins": 4.838027000427246,
+      "rewards/rejected": -11.024351119995117,
+      "step": 690
+    },
+    {
+      "epoch": 0.746268656716418,
+      "grad_norm": 4.507752895355225,
+      "learning_rate": 2.761321158169134e-06,
+      "logits/chosen": 2.295815944671631,
+      "logits/rejected": 1.4525251388549805,
+      "logps/chosen": -3.4264607429504395,
+      "logps/rejected": -6.397761344909668,
+      "loss": 0.3257,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -5.139691352844238,
+      "rewards/margins": 4.45695161819458,
+      "rewards/rejected": -9.596643447875977,
+      "step": 700
+    },
+    {
+      "epoch": 0.746268656716418,
+      "eval_logits/chosen": 1.9169458150863647,
+      "eval_logits/rejected": 1.761539101600647,
+      "eval_logps/chosen": -3.324993848800659,
+      "eval_logps/rejected": -6.471699237823486,
+      "eval_loss": 0.3062504529953003,
+      "eval_rewards/accuracies": 0.9473684430122375,
+      "eval_rewards/chosen": -4.987491607666016,
+      "eval_rewards/margins": 4.720058441162109,
+      "eval_rewards/rejected": -9.707548141479492,
+      "eval_runtime": 21.4456,
+      "eval_samples_per_second": 28.304,
+      "eval_steps_per_second": 3.544,
+      "step": 700
+    },
+    {
+      "epoch": 0.7569296375266524,
+      "grad_norm": 6.864930629730225,
+      "learning_rate": 2.70919460833079e-06,
+      "logits/chosen": 1.8696104288101196,
+      "logits/rejected": 1.4743572473526,
+      "logps/chosen": -3.4536032676696777,
+      "logps/rejected": -6.400083065032959,
+      "loss": 0.3255,
+      "rewards/accuracies": 0.8999999761581421,
+      "rewards/chosen": -5.180405616760254,
+      "rewards/margins": 4.4197187423706055,
+      "rewards/rejected": -9.60012435913086,
+      "step": 710
+    },
+    {
+      "epoch": 0.767590618336887,
+      "grad_norm": 5.012319087982178,
+      "learning_rate": 2.6569762988232838e-06,
+      "logits/chosen": 2.861811399459839,
+      "logits/rejected": 2.266331911087036,
+      "logps/chosen": -4.254661560058594,
+      "logps/rejected": -6.501154899597168,
+      "loss": 0.2845,
+      "rewards/accuracies": 0.8999999761581421,
+      "rewards/chosen": -6.381992816925049,
+      "rewards/margins": 3.3697407245635986,
+      "rewards/rejected": -9.751731872558594,
+      "step": 720
+    },
+    {
+      "epoch": 0.7782515991471215,
+      "grad_norm": 1.460620403289795,
+      "learning_rate": 2.604689134322999e-06,
+      "logits/chosen": 2.917189836502075,
+      "logits/rejected": 2.3850274085998535,
+      "logps/chosen": -3.6072356700897217,
+      "logps/rejected": -6.3972039222717285,
+      "loss": 0.2696,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -5.410854339599609,
+      "rewards/margins": 4.184950828552246,
+      "rewards/rejected": -9.595805168151855,
+      "step": 730
+    },
+    {
+      "epoch": 0.7889125799573561,
+      "grad_norm": 5.478528022766113,
+      "learning_rate": 2.5523560497083927e-06,
+      "logits/chosen": 2.2212367057800293,
+      "logits/rejected": 1.6661183834075928,
+      "logps/chosen": -3.353309154510498,
+      "logps/rejected": -6.527769565582275,
+      "loss": 0.2622,
+      "rewards/accuracies": 0.949999988079071,
+      "rewards/chosen": -5.029964447021484,
+      "rewards/margins": 4.761690616607666,
+      "rewards/rejected": -9.791654586791992,
+      "step": 740
+    },
+    {
+      "epoch": 0.7995735607675906,
+      "grad_norm": 2.7094156742095947,
+      "learning_rate": 2.5e-06,
+      "logits/chosen": 2.9158451557159424,
+      "logits/rejected": 2.286925792694092,
+      "logps/chosen": -4.0971856117248535,
+      "logps/rejected": -6.912562370300293,
+      "loss": 0.3243,
+      "rewards/accuracies": 0.887499988079071,
+      "rewards/chosen": -6.145778179168701,
+      "rewards/margins": 4.2230658531188965,
+      "rewards/rejected": -10.368844985961914,
+      "step": 750
+    },
+    {
+      "epoch": 0.7995735607675906,
+      "eval_logits/chosen": 2.0261385440826416,
+      "eval_logits/rejected": 1.8928862810134888,
+      "eval_logps/chosen": -3.538414478302002,
+      "eval_logps/rejected": -6.846959114074707,
+      "eval_loss": 0.2867603600025177,
+      "eval_rewards/accuracies": 0.9342105388641357,
+      "eval_rewards/chosen": -5.307621479034424,
+      "eval_rewards/margins": 4.962815761566162,
+      "eval_rewards/rejected": -10.270438194274902,
+      "eval_runtime": 21.4446,
+      "eval_samples_per_second": 28.305,
+      "eval_steps_per_second": 3.544,
+      "step": 750
+    },
+    {
+      "epoch": 0.8102345415778252,
+      "grad_norm": 2.477257490158081,
+      "learning_rate": 2.447643950291608e-06,
+      "logits/chosen": 1.8020350933074951,
+      "logits/rejected": 1.3997737169265747,
+      "logps/chosen": -3.34975004196167,
+      "logps/rejected": -6.07710075378418,
+      "loss": 0.2394,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -5.024625301361084,
+      "rewards/margins": 4.091025352478027,
+      "rewards/rejected": -9.11565113067627,
+      "step": 760
+    },
+    {
+      "epoch": 0.8208955223880597,
+      "grad_norm": 4.726482391357422,
+      "learning_rate": 2.3953108656770018e-06,
+      "logits/chosen": 2.757497787475586,
+      "logits/rejected": 2.435739517211914,
+      "logps/chosen": -4.294445991516113,
+      "logps/rejected": -7.038791656494141,
+      "loss": 0.3113,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -6.441669464111328,
+      "rewards/margins": 4.116518497467041,
+      "rewards/rejected": -10.558187484741211,
+      "step": 770
+    },
+    {
+      "epoch": 0.8315565031982942,
+      "grad_norm": 2.759181022644043,
+      "learning_rate": 2.3430237011767166e-06,
+      "logits/chosen": 2.788053274154663,
+      "logits/rejected": 2.1961398124694824,
+      "logps/chosen": -3.792672634124756,
+      "logps/rejected": -7.1101579666137695,
+      "loss": 0.3043,
+      "rewards/accuracies": 0.9125000238418579,
+      "rewards/chosen": -5.689009666442871,
+      "rewards/margins": 4.976227760314941,
+      "rewards/rejected": -10.665237426757812,
+      "step": 780
+    },
+    {
+      "epoch": 0.8422174840085288,
+      "grad_norm": 3.382880687713623,
+      "learning_rate": 2.290805391669212e-06,
+      "logits/chosen": 2.3644931316375732,
+      "logits/rejected": 1.838971495628357,
+      "logps/chosen": -3.919532060623169,
+      "logps/rejected": -7.125067710876465,
+      "loss": 0.2825,
+      "rewards/accuracies": 0.9375,
+      "rewards/chosen": -5.879298210144043,
+      "rewards/margins": 4.8083038330078125,
+      "rewards/rejected": -10.687601089477539,
+      "step": 790
+    },
+    {
+      "epoch": 0.8528784648187633,
+      "grad_norm": 3.331052541732788,
+      "learning_rate": 2.238678841830867e-06,
+      "logits/chosen": 2.3921093940734863,
+      "logits/rejected": 1.9997615814208984,
+      "logps/chosen": -3.46040678024292,
+      "logps/rejected": -6.8662567138671875,
+      "loss": 0.3094,
+      "rewards/accuracies": 0.949999988079071,
+      "rewards/chosen": -5.190610885620117,
+      "rewards/margins": 5.108774662017822,
+      "rewards/rejected": -10.299385070800781,
+      "step": 800
+    },
+    {
+      "epoch": 0.8528784648187633,
+      "eval_logits/chosen": 2.198673963546753,
+      "eval_logits/rejected": 2.091115713119507,
+      "eval_logps/chosen": -3.606259346008301,
+      "eval_logps/rejected": -7.107792854309082,
+      "eval_loss": 0.27349653840065,
+      "eval_rewards/accuracies": 0.9473684430122375,
+      "eval_rewards/chosen": -5.409388542175293,
+      "eval_rewards/margins": 5.25230073928833,
+      "eval_rewards/rejected": -10.661689758300781,
+      "eval_runtime": 21.4427,
+      "eval_samples_per_second": 28.308,
+      "eval_steps_per_second": 3.544,
+      "step": 800
+    },
+    {
+      "epoch": 0.8635394456289979,
+      "grad_norm": 1.6861388683319092,
+      "learning_rate": 2.186666916089239e-06,
+      "logits/chosen": 3.2604496479034424,
+      "logits/rejected": 2.558086395263672,
+      "logps/chosen": -3.778765916824341,
+      "logps/rejected": -6.954268455505371,
+      "loss": 0.2923,
+      "rewards/accuracies": 0.925000011920929,
+      "rewards/chosen": -5.668148994445801,
+      "rewards/margins": 4.763254165649414,
+      "rewards/rejected": -10.431403160095215,
+      "step": 810
+    },
+    {
+      "epoch": 0.8742004264392325,
+      "grad_norm": 3.100459575653076,
+      "learning_rate": 2.134792428593971e-06,
+      "logits/chosen": 2.6463425159454346,
+      "logits/rejected": 2.2412638664245605,
+      "logps/chosen": -4.3968634605407715,
+      "logps/rejected": -7.397219181060791,
+      "loss": 0.2767,
+      "rewards/accuracies": 0.862500011920929,
+      "rewards/chosen": -6.595294952392578,
+      "rewards/margins": 4.500533103942871,
+      "rewards/rejected": -11.09582805633545,
+      "step": 820
+    },
+    {
+      "epoch": 0.8848614072494669,
+      "grad_norm": 2.7398672103881836,
+      "learning_rate": 2.0830781332097446e-06,
+      "logits/chosen": 2.8484559059143066,
+      "logits/rejected": 2.408825635910034,
+      "logps/chosen": -3.5274035930633545,
+      "logps/rejected": -6.786914825439453,
+      "loss": 0.2587,
+      "rewards/accuracies": 0.8999999761581421,
+      "rewards/chosen": -5.291105270385742,
+      "rewards/margins": 4.889266014099121,
+      "rewards/rejected": -10.180373191833496,
+      "step": 830
+    },
+    {
+      "epoch": 0.8955223880597015,
+      "grad_norm": 4.221662521362305,
+      "learning_rate": 2.031546713535688e-06,
+      "logits/chosen": 2.5181362628936768,
+      "logits/rejected": 2.1050801277160645,
+      "logps/chosen": -3.7997944355010986,
+      "logps/rejected": -7.4044928550720215,
+      "loss": 0.2985,
+      "rewards/accuracies": 0.9750000238418579,
+      "rewards/chosen": -5.6996917724609375,
+      "rewards/margins": 5.407048225402832,
+      "rewards/rejected": -11.10673999786377,
+      "step": 840
+    },
+    {
+      "epoch": 0.906183368869936,
+      "grad_norm": 4.389741897583008,
+      "learning_rate": 1.9802207729556023e-06,
+      "logits/chosen": 1.7594547271728516,
+      "logits/rejected": 1.5125747919082642,
+      "logps/chosen": -3.7186477184295654,
+      "logps/rejected": -7.241146087646484,
+      "loss": 0.278,
+      "rewards/accuracies": 0.949999988079071,
+      "rewards/chosen": -5.577971935272217,
+      "rewards/margins": 5.28374719619751,
+      "rewards/rejected": -10.861719131469727,
+      "step": 850
+    },
+    {
+      "epoch": 0.906183368869936,
+      "eval_logits/chosen": 2.196138381958008,
+      "eval_logits/rejected": 2.091784715652466,
+      "eval_logps/chosen": -3.610405683517456,
+      "eval_logps/rejected": -7.241917610168457,
+      "eval_loss": 0.2656216025352478,
+      "eval_rewards/accuracies": 0.9473684430122375,
+      "eval_rewards/chosen": -5.415609359741211,
+      "eval_rewards/margins": 5.447267532348633,
+      "eval_rewards/rejected": -10.862876892089844,
+      "eval_runtime": 21.4468,
+      "eval_samples_per_second": 28.303,
+      "eval_steps_per_second": 3.544,
+      "step": 850
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 1500,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 50,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.0364770657288847e+18,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-850/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29030fd2c9579a53d4cbb5166f67721176cc7b02c89ae9a2d5722caa76c57c96
+size 7224

checkpoint-850/zero_to_fp32.py

@@ -0,0 +1,674 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import json
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering Sharded Weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        state_dict_split = split_torch_state_dict_into_shards(state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard = {tensor: state_dict[tensor].contiguous() for tensor in tensors}
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard, output_path)
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)