423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-04-27 18:29:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

[data] support discard history cot for multiturn (#10435)

Browse Source
This commit is contained in:
Kingsley
2026-04-27 00:32:44 +08:00
committed by GitHub
parent 79c8332e4c
commit c8890c32db
3 changed files with 47 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
tests/data/test_template.py
View File

@@ -181,6 +181,39 @@ def test_reasoning_encode_multiturn(cot_messages: bool, enable_thinking: bool):
(prompt_str_1, answer_str_1, prompt_str_2, answer_str_2),
)
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("enable_thinking", [True, False, None])
@pytest.mark.parametrize("discarding_history_cot", [True, False])
def test_reasoning_encode_multiturn_discarding_history_cot(enable_thinking: bool, discarding_history_cot: bool):
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
data_args = DataArguments(template="qwen3", enable_thinking=enable_thinking)
template = get_template_and_fix_tokenizer(tokenizer, data_args)
encoded_pairs = template.encode_multiturn(tokenizer, MESSAGES_WITH_THOUGHT, discarding_history_cot=discarding_history_cot)
prompt_str_1 = f"<|im_start|>user\n{MESSAGES_WITH_THOUGHT[0]['content']}<|im_end|>\n<|im_start|>assistant\n"
prompt_str_2 = f"<|im_start|>user\n{MESSAGES_WITH_THOUGHT[2]['content']}<|im_end|>\n<|im_start|>assistant\n"
if enable_thinking is False:
answer_str_1 = f"{MESSAGES[1]['content']}<|im_end|>\n"
answer_str_2 = f"{MESSAGES[3]['content']}<|im_end|>\n"
if discarding_history_cot:
prompt_str_2 = prompt_str_2 + "<think>\n\n</think>\n\n"
else:
prompt_str_1 = prompt_str_1 + "<think>\n\n</think>\n\n"
prompt_str_2 = prompt_str_2 + "<think>\n\n</think>\n\n"
else:
if discarding_history_cot:
answer_str_1 = f"{MESSAGES[1]['content']}<|im_end|>\n"
else:
answer_str_1 = f"{MESSAGES_WITH_THOUGHT[1]['content']}<|im_end|>\n"
answer_str_2 = f"{MESSAGES_WITH_THOUGHT[3]['content']}<|im_end|>\n"
_check_tokenization(
tokenizer,
(encoded_pairs[0][0], encoded_pairs[0][1], encoded_pairs[1][0], encoded_pairs[1][1]),
(prompt_str_1, answer_str_1, prompt_str_2, answer_str_2),
)
@pytest.mark.runs_on(["cpu", "mps"])
def test_jinja_template():