fix new dataset prompt tokenizers
Browse files
src/axolotl/datasets.py
CHANGED
|
@@ -106,7 +106,7 @@ class ConstantLengthDataset(IterableDataset):
|
|
| 106 |
}
|
| 107 |
else:
|
| 108 |
logging.warning(
|
| 109 |
-
"dropping batch due to tensor size mismatch"
|
| 110 |
)
|
| 111 |
buffer = {"input_ids": [], "attention_mask": [], "labels": []}
|
| 112 |
buffer_len = 0
|
|
|
|
| 106 |
}
|
| 107 |
else:
|
| 108 |
logging.warning(
|
| 109 |
+
f"dropping batch due to tensor size mismatch input_ids: {input_ids.size()}, labels: {labels.size()}, attention_mask: {attention_mask.size()}"
|
| 110 |
)
|
| 111 |
buffer = {"input_ids": [], "attention_mask": [], "labels": []}
|
| 112 |
buffer_len = 0
|
src/axolotl/prompt_strategies/__init__.py
CHANGED
|
@@ -1,11 +1,13 @@
|
|
| 1 |
import importlib
|
| 2 |
-
from functools import cache
|
| 3 |
|
| 4 |
-
@cache
|
| 5 |
def load(strategy, tokenizer, cfg):
|
| 6 |
try:
|
| 7 |
-
|
| 8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
return fn(tokenizer, cfg)
|
| 10 |
except:
|
| 11 |
pass
|
|
|
|
| 1 |
import importlib
|
|
|
|
| 2 |
|
|
|
|
| 3 |
def load(strategy, tokenizer, cfg):
|
| 4 |
try:
|
| 5 |
+
load_fn = "load"
|
| 6 |
+
if strategy.split(".")[-1].startswith("load_"):
|
| 7 |
+
load_fn = strategy.split(".")[-1]
|
| 8 |
+
strategy = ".".join(strategy.split(".")[:-1])
|
| 9 |
+
m = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies")
|
| 10 |
+
fn = getattr(m, load_fn)
|
| 11 |
return fn(tokenizer, cfg)
|
| 12 |
except:
|
| 13 |
pass
|
src/axolotl/prompt_strategies/creative_acr.py
ADDED
|
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Union, Generator
|
| 2 |
+
|
| 3 |
+
import yaml
|
| 4 |
+
from axolotl.prompt_tokenizers import InstructionPromptTokenizingStrategy
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
class CreativeAnsweringPromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
| 8 |
+
def parse_instruction_fields(self, prompt) -> (str, str, str):
|
| 9 |
+
question = prompt["instruction"]
|
| 10 |
+
answer = prompt["revision"] # don't use prompt[answer], that's data we don't want in the dataset
|
| 11 |
+
return (
|
| 12 |
+
question,
|
| 13 |
+
"",
|
| 14 |
+
answer,
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class CreativeCritiquePromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
| 19 |
+
user_prompt = """Given the following Question and Response, critique the Response on a scale of 1-10. You should critique the answer in the following criteria:
|
| 20 |
+
refusal: whether the Response refuses to answer the Question. Responses containing language similar to "As an AI language model, I am not capable of ..." is a form of refusal and should have a low refusal score. 10 is no refusal, 0 is refuses to answer the question.
|
| 21 |
+
prescriptive bias: whether the Response attempts to prescribe or dictate certain actions or behaviors of the user. 10 no prescriptive bias present, 0 means there is prescriptive bias.
|
| 22 |
+
creativity: score how creative the Response is. 10 is most creative, 0 lacks creativity
|
| 23 |
+
comprehensiveness: score how comprehensive and in-depth the Response is. 10 is most comprehensive and in-depth, 0 lacks coverage and depth
|
| 24 |
+
Then for each criteria, give a detailed explanation of why the response doesn't meet the criteria (don't reference the score), what it did right, and how it could be improved for that criteria.
|
| 25 |
+
|
| 26 |
+
Output your answer in YAML format like so:
|
| 27 |
+
scores:
|
| 28 |
+
refusal: <0-10>
|
| 29 |
+
prescriptive_bias: <0-10>
|
| 30 |
+
creativity: <0-10>
|
| 31 |
+
comprehensiveness: <0-10>
|
| 32 |
+
critiques:
|
| 33 |
+
refusal:
|
| 34 |
+
explanation: ...
|
| 35 |
+
improvements: ...
|
| 36 |
+
prescriptive_bias:
|
| 37 |
+
explanation: ...
|
| 38 |
+
improvements: ...
|
| 39 |
+
creativity:
|
| 40 |
+
explanation: ...
|
| 41 |
+
improvements: ...
|
| 42 |
+
comprehensiveness:
|
| 43 |
+
explanation: ...
|
| 44 |
+
improvements: ...
|
| 45 |
+
|
| 46 |
+
Question: {question}
|
| 47 |
+
Answer: {answer}
|
| 48 |
+
"""
|
| 49 |
+
|
| 50 |
+
def parse_instruction_fields(self, prompt) -> (str, str, str):
|
| 51 |
+
scores = yaml.dump(prompt["scores"], default_flow_style=False, Dumper=yaml.Dumper)
|
| 52 |
+
critiques = yaml.dump(prompt["critiques"], default_flow_style=False, Dumper=yaml.Dumper)
|
| 53 |
+
evaluation = scores + critiques
|
| 54 |
+
question = prompt["instruction"]
|
| 55 |
+
answer = prompt["answer"]
|
| 56 |
+
return (
|
| 57 |
+
self.user_prompt.format(question=question, answer=answer),
|
| 58 |
+
"",
|
| 59 |
+
evaluation,
|
| 60 |
+
)
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
class CreativeRevisePromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
| 64 |
+
user_prompt = """Definitions:
|
| 65 |
+
refusal: whether the Response refuses to answer the Question. Responses containing language similar to "As an AI language model, I am not capable of ..." is a form of refusal and should have a low refusal score. 10 is no refusal, 0 is refuses to answer the question.
|
| 66 |
+
prescriptive bias: whether the Response attempts to prescribe or dictate certain actions or behaviors of the user. 10 no prescriptive bias present, 0 means their is prescriptive bias.
|
| 67 |
+
creativity: score how creative the Response is. 10 is most creative, 0 lacks creativity
|
| 68 |
+
comprehensiveness: score how comprehensive and in-depth the Response is. 10 is most comprehensive and in-depth, 0 lacks coverage and depth
|
| 69 |
+
|
| 70 |
+
Given the following Question, Response, and Evaluation, revise the Response based on the Evaluation and recommendations for improvements. Reply only with the revised response.
|
| 71 |
+
|
| 72 |
+
Question: {question}
|
| 73 |
+
Answer: {answer}
|
| 74 |
+
Evaluation:
|
| 75 |
+
{evaluation}
|
| 76 |
+
"""
|
| 77 |
+
|
| 78 |
+
def parse_instruction_fields(self, prompt) -> (str, str, str):
|
| 79 |
+
scores = yaml.dump(prompt["scores"], default_flow_style=False, Dumper=yaml.Dumper)
|
| 80 |
+
critiques = yaml.dump(prompt["critiques"], default_flow_style=False, Dumper=yaml.Dumper)
|
| 81 |
+
evaluation = scores + critiques
|
| 82 |
+
question = prompt["instruction"]
|
| 83 |
+
answer = prompt["answer"]
|
| 84 |
+
return (
|
| 85 |
+
self.user_prompt.format(question=question, answer=answer, evaluation=evaluation),
|
| 86 |
+
"",
|
| 87 |
+
prompt["revision"],
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
class CreativePrompterBase:
|
| 92 |
+
system_prompt = ""
|
| 93 |
+
prompt_input = "{system_prompt}\nUSER: {instruction}\nASSISTANT:"
|
| 94 |
+
|
| 95 |
+
def build_prompt(
|
| 96 |
+
self,
|
| 97 |
+
instruction: str,
|
| 98 |
+
input: Union[None, str] = None,
|
| 99 |
+
output: Union[None, str] = None,
|
| 100 |
+
) -> Generator[str, None, None]:
|
| 101 |
+
if self.system_prompt:
|
| 102 |
+
res = f"{self.system_prompt}\nUSER: {instruction}\nASSISTANT:"
|
| 103 |
+
else:
|
| 104 |
+
res = f"USER: {instruction}\nASSISTANT:"
|
| 105 |
+
if output:
|
| 106 |
+
res = f"{res}{output}"
|
| 107 |
+
yield res
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
class CreativeAnswerPrompter(CreativePrompterBase):
|
| 111 |
+
system_prompt = "Answer the following question in a comprehensive, in-depth, and creative way. Additionally your response should be relevant, accurate, and free of any ambiguity."
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
class CreativeCritiquePrompter(CreativePrompterBase):
|
| 115 |
+
system_prompt = ""
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
class CreativeRevisePrompter(CreativePrompterBase):
|
| 119 |
+
system_prompt = ""
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def load_answer(tokenizer, cfg):
|
| 123 |
+
return CreativeAnsweringPromptTokenizingStrategy(
|
| 124 |
+
CreativeAnswerPrompter(), tokenizer, cfg.train_on_inputs, cfg.sequence_len
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
|
| 128 |
+
def load_critique(tokenizer, cfg):
|
| 129 |
+
return CreativeCritiquePromptTokenizingStrategy(
|
| 130 |
+
CreativeCritiquePrompter(), tokenizer, cfg.train_on_inputs, cfg.sequence_len
|
| 131 |
+
)
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
def load_revise(tokenizer, cfg):
|
| 135 |
+
return CreativeRevisePromptTokenizingStrategy(
|
| 136 |
+
CreativeRevisePrompter(), tokenizer, cfg.train_on_inputs, cfg.sequence_len
|
| 137 |
+
)
|
src/axolotl/prompt_strategies/pygmalion.py
CHANGED
|
@@ -41,9 +41,9 @@ class PygmalionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
| 41 |
elif role == "bot":
|
| 42 |
prefix = "<|model|>"
|
| 43 |
res = self._tokenize(prefix + " " + message.strip(), add_eos_token=True, strip_bos_token=True)
|
| 44 |
-
res["input_ids"] = [*self.bot_prefix_token_ids, *res["input_ids"]]
|
| 45 |
# mask out the prefix token, rest is not masked out from labels
|
| 46 |
-
|
|
|
|
| 47 |
else:
|
| 48 |
logging.warning(f"unknown role in conversation: {role}")
|
| 49 |
res = defaultdict(lambda: [])
|
|
|
|
| 41 |
elif role == "bot":
|
| 42 |
prefix = "<|model|>"
|
| 43 |
res = self._tokenize(prefix + " " + message.strip(), add_eos_token=True, strip_bos_token=True)
|
|
|
|
| 44 |
# mask out the prefix token, rest is not masked out from labels
|
| 45 |
+
# make sure we create the labels first, otherwise we get incorrect lengths
|
| 46 |
+
labels = [ IGNORE_TOKEN_ID ] * len(self.bot_prefix_token_ids) + [*copy.deepcopy(res["input_ids"])][len(self.bot_prefix_token_ids):]
|
| 47 |
else:
|
| 48 |
logging.warning(f"unknown role in conversation: {role}")
|
| 49 |
res = defaultdict(lambda: [])
|
src/axolotl/utils/data.py
CHANGED
|
@@ -75,7 +75,7 @@ def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_pa
|
|
| 75 |
ds = None
|
| 76 |
ds_from_hub = False
|
| 77 |
try:
|
| 78 |
-
load_dataset(d.path, streaming=True)
|
| 79 |
ds_from_hub = True
|
| 80 |
except FileNotFoundError:
|
| 81 |
pass
|
|
@@ -83,18 +83,18 @@ def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_pa
|
|
| 83 |
# prefer local dataset, even if hub exists
|
| 84 |
if Path(d.path).exists():
|
| 85 |
ds: IterableDataset = load_dataset(
|
| 86 |
-
"json", data_files=d.path, streaming=
|
| 87 |
)
|
| 88 |
elif ds_from_hub:
|
| 89 |
if d.data_files:
|
| 90 |
-
ds = load_dataset(d.path, streaming=
|
| 91 |
else:
|
| 92 |
-
ds = load_dataset(d.path, streaming=True)
|
| 93 |
else:
|
| 94 |
fp = hf_hub_download(
|
| 95 |
repo_id=d.path, repo_type="dataset", filename=d.data_files
|
| 96 |
)
|
| 97 |
-
ds = load_dataset("json", data_files=fp, streaming=
|
| 98 |
if not ds:
|
| 99 |
raise Exception("unhandled dataset load")
|
| 100 |
d_type = d.type
|
|
|
|
| 75 |
ds = None
|
| 76 |
ds_from_hub = False
|
| 77 |
try:
|
| 78 |
+
load_dataset(d.path, streaming=True, use_auth_token=True)
|
| 79 |
ds_from_hub = True
|
| 80 |
except FileNotFoundError:
|
| 81 |
pass
|
|
|
|
| 83 |
# prefer local dataset, even if hub exists
|
| 84 |
if Path(d.path).exists():
|
| 85 |
ds: IterableDataset = load_dataset(
|
| 86 |
+
"json", data_files=d.path, streaming=False, split=None
|
| 87 |
)
|
| 88 |
elif ds_from_hub:
|
| 89 |
if d.data_files:
|
| 90 |
+
ds = load_dataset(d.path, streaming=False, data_files=d.data_files, use_auth_token=True)
|
| 91 |
else:
|
| 92 |
+
ds = load_dataset(d.path, streaming=False, use_auth_token=True)
|
| 93 |
else:
|
| 94 |
fp = hf_hub_download(
|
| 95 |
repo_id=d.path, repo_type="dataset", filename=d.data_files
|
| 96 |
)
|
| 97 |
+
ds = load_dataset("json", data_files=fp, streaming=False, split=None)
|
| 98 |
if not ds:
|
| 99 |
raise Exception("unhandled dataset load")
|
| 100 |
d_type = d.type
|