refactor: SFT 统一 messages 格式 + ChatML 纯 jinja 渲染
This commit is contained in:
+22
-38
@@ -7,7 +7,7 @@ from torch import Tensor
|
||||
|
||||
from pipeline.tokenize import AutoTokenizer
|
||||
from pipeline.strategies import PromptStrategy, ChatMLStrategy
|
||||
from pipeline.processors.base import BaseProcessor, ProcessorSchema, encode_with_mask
|
||||
from pipeline.processors.base import BaseProcessor, ProcessorSchema
|
||||
from pipeline.processors.factory import ProcessorFactory
|
||||
|
||||
|
||||
@@ -15,20 +15,20 @@ from pipeline.processors.factory import ProcessorFactory
|
||||
class SFTProcessor(BaseProcessor):
|
||||
"""Supervised fine-tuning data processor.
|
||||
|
||||
Supports two input formats:
|
||||
Input formats:
|
||||
1. messages (recommended):
|
||||
``{"messages": [{"role": "user", "content": "..."},
|
||||
{"role": "assistant", "content": "..."}]}``
|
||||
Multi-turn and system prompts are supported.
|
||||
The tokenizer's ``apply_chat_template`` is used for rendering.
|
||||
Multi-turn and system prompts are supported. Each assistant
|
||||
turn gets ``loss_mask = 1``; all other roles get 0.
|
||||
2. legacy query/response:
|
||||
``{"query": "...", "response": "..."}``
|
||||
Falls back to the configured PromptStrategy (ChatML by default).
|
||||
Internally converted to messages.
|
||||
|
||||
Output schema:
|
||||
- sequence: int32 tensor - Combined token IDs (prompt + response)
|
||||
- loss_mask: bool tensor - True for response tokens (compute loss)
|
||||
- position_ids: int32 tensor - Per-sample position IDs starting from 0
|
||||
- sequence: int32 tensor - Combined token IDs
|
||||
- loss_mask: bool tensor - True for assistant response tokens
|
||||
- position_ids: int32 tensor - Per-sample position IDs, start from 0
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
@@ -58,7 +58,10 @@ class SFTProcessor(BaseProcessor):
|
||||
if "messages" in input_dict:
|
||||
return self._process_messages(input_dict["messages"])
|
||||
if "query" in input_dict and "response" in input_dict:
|
||||
return self._process_legacy(input_dict)
|
||||
return self._process_messages([
|
||||
{"role": "user", "content": input_dict["query"]},
|
||||
{"role": "assistant", "content": input_dict["response"]},
|
||||
])
|
||||
raise KeyError(
|
||||
"Input must contain 'messages' or 'query'/'response' pair"
|
||||
)
|
||||
@@ -69,38 +72,19 @@ class SFTProcessor(BaseProcessor):
|
||||
if messages[-1]["role"] != "assistant":
|
||||
raise ValueError("Last message must have role 'assistant'")
|
||||
|
||||
last_asst_idx = max(
|
||||
i for i, m in enumerate(messages) if m["role"] == "assistant"
|
||||
)
|
||||
|
||||
prompt_tokens = self.tokenizer.apply_chat_template(
|
||||
messages[:last_asst_idx],
|
||||
add_generation_prompt=True,
|
||||
tokenize=True,
|
||||
)
|
||||
|
||||
resp_content = messages[last_asst_idx]["content"]
|
||||
im_end = getattr(self.tokenizer, "im_end", "<|im_end|>")
|
||||
resp_tokens = self.tokenizer.encode(
|
||||
f"{resp_content}{im_end}\n", add_special_tokens=False
|
||||
)
|
||||
|
||||
tokens, loss_mask = encode_with_mask(prompt_tokens, resp_tokens)
|
||||
position_ids = torch.arange(len(tokens), dtype=torch.int32)
|
||||
return {"sequence": tokens, "loss_mask": loss_mask, "position_ids": position_ids}
|
||||
|
||||
def _process_legacy(self, input_dict: Dict[str, Any]) -> Dict[str, Tensor]:
|
||||
strategy = self.strategy or ChatMLStrategy(self.tokenizer)
|
||||
|
||||
query_tokens = self.tokenizer.encode(input_dict["query"])
|
||||
response_tokens = self.tokenizer.encode(input_dict["response"])
|
||||
prompt, resp = strategy.format_messages(messages)
|
||||
|
||||
prompt = strategy.assemble_prompt(query_tokens)
|
||||
response = strategy.assemble_response(response_tokens)
|
||||
|
||||
tokens, loss_mask = encode_with_mask(prompt, response)
|
||||
position_ids = torch.arange(len(tokens), dtype=torch.int32)
|
||||
return {"sequence": tokens, "loss_mask": loss_mask, "position_ids": position_ids}
|
||||
sequence = torch.tensor(prompt + resp, dtype=torch.int32)
|
||||
loss_mask = torch.zeros(len(sequence), dtype=torch.bool)
|
||||
loss_mask[len(prompt) :] = True
|
||||
position_ids = torch.arange(len(sequence), dtype=torch.int32)
|
||||
return {
|
||||
"sequence": sequence,
|
||||
"loss_mask": loss_mask,
|
||||
"position_ids": position_ids,
|
||||
}
|
||||
|
||||
@property
|
||||
def output_keys(self) -> List[str]:
|
||||
|
||||
@@ -1,43 +1,91 @@
|
||||
"""ChatML format strategy."""
|
||||
|
||||
from typing import List
|
||||
from typing import Dict, List, Tuple
|
||||
|
||||
from pipeline.tokenize import AutoTokenizer
|
||||
from pipeline.strategies.base import PromptStrategy
|
||||
from pipeline.strategies.factory import StrategyFactory
|
||||
|
||||
DEFAULT_CHATML_TEMPLATE = (
|
||||
"{% for message in messages %}"
|
||||
"{% if message['role'] == 'system' %}"
|
||||
"{{ '<|im_start|>system\n' + message['content'] + '<|im_end|>\n' }}"
|
||||
"{% elif message['role'] == 'user' %}"
|
||||
"{{ '<|im_start|>user\n' + message['content'] + '<|im_end|>\n' }}"
|
||||
"{% elif message['role'] == 'assistant' %}"
|
||||
"{{ '<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' }}"
|
||||
"{% endif %}"
|
||||
"{% endfor %}"
|
||||
"{% if add_generation_prompt %}"
|
||||
"{{ '<|im_start|>assistant\n' }}"
|
||||
"{% endif %}"
|
||||
)
|
||||
|
||||
|
||||
@StrategyFactory.register("chatml")
|
||||
class ChatMLStrategy(PromptStrategy):
|
||||
"""ChatML format strategy."""
|
||||
"""ChatML format strategy.
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
tokenizer: AutoTokenizer,
|
||||
user_start: str = "<｜im▁start｜>user",
|
||||
user_end: str = "<｜im▁end｜>",
|
||||
assistant_start: str = "<｜im▁start｜>assistant",
|
||||
assistant_end: str = "<｜im▁end｜>",
|
||||
):
|
||||
Renders messages using the tokenizer's jinja chat_template from
|
||||
``tokenizer_config.json``. Falls back to DEFAULT_CHATML_TEMPLATE
|
||||
when no template is configured.
|
||||
|
||||
The strategy does **not** hard-code any special tokens – all
|
||||
formatting is driven by the jinja template.
|
||||
"""
|
||||
|
||||
def __init__(self, tokenizer: AutoTokenizer):
|
||||
super().__init__(tokenizer)
|
||||
nl_id = tokenizer.encode("a\nb", add_special_tokens=False)[1]
|
||||
|
||||
self._user_start_ids = self._encode_format(user_start) + [nl_id]
|
||||
self._user_end_ids = self._encode_format(user_end) + [nl_id]
|
||||
self._assistant_start_ids = self._encode_format(assistant_start) + [nl_id]
|
||||
self._assistant_end_ids = self._encode_format(assistant_end) + [nl_id]
|
||||
if tokenizer._chat_template is None:
|
||||
tokenizer.set_chat_template(DEFAULT_CHATML_TEMPLATE)
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
return "chatml"
|
||||
|
||||
def format_messages(
|
||||
self,
|
||||
messages: List[Dict[str, str]],
|
||||
) -> Tuple[List[int], List[int]]:
|
||||
"""Render a single-turn messages conversation.
|
||||
|
||||
Returns ``(prompt_tokens, response_tokens)`` where
|
||||
*prompt_tokens* contains everything up to (and including) the
|
||||
last assistant start marker, and *response_tokens* is the
|
||||
assistant content plus the closing markers.
|
||||
"""
|
||||
last_asst = max(
|
||||
i for i, m in enumerate(messages) if m["role"] == "assistant"
|
||||
)
|
||||
|
||||
prompt = self.tokenizer.apply_chat_template(
|
||||
messages[:last_asst],
|
||||
add_generation_prompt=True,
|
||||
tokenize=True,
|
||||
)
|
||||
full = self.tokenizer.apply_chat_template(
|
||||
messages[: last_asst + 1],
|
||||
add_generation_prompt=False,
|
||||
tokenize=True,
|
||||
)
|
||||
return prompt, full[len(prompt) :]
|
||||
|
||||
def assemble_prompt(self, query_tokens: List[int]) -> List[int]:
|
||||
return (
|
||||
self._user_start_ids
|
||||
+ query_tokens
|
||||
+ self._user_end_ids
|
||||
+ self._assistant_start_ids
|
||||
text = self.tokenizer.decode(query_tokens)
|
||||
return self.tokenizer.apply_chat_template(
|
||||
[{"role": "user", "content": text}],
|
||||
add_generation_prompt=True,
|
||||
tokenize=True,
|
||||
)
|
||||
|
||||
def assemble_response(self, response_tokens: List[int]) -> List[int]:
|
||||
return response_tokens + self._assistant_end_ids
|
||||
text = self.tokenizer.decode(response_tokens)
|
||||
full = self.tokenizer.apply_chat_template(
|
||||
[{"role": "assistant", "content": text}],
|
||||
add_generation_prompt=False,
|
||||
tokenize=True,
|
||||
)
|
||||
opening = self.tokenizer.apply_chat_template(
|
||||
[], add_generation_prompt=True, tokenize=True
|
||||
)
|
||||
return full[len(opening) :]
|
||||
|
||||
Reference in New Issue
Block a user