423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-01-13 17:40:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v1] add sft (#9752)

Browse Source
This commit is contained in:
Yaowei Zheng
2026-01-12 03:15:01 +08:00
committed by GitHub
parent 4d3621e3d3
commit 958b9c3468
29 changed files with 439 additions and 305 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests_v1/core/test_data_engine.py
View File

@@ -23,8 +23,8 @@ from llamafactory.v1.core.data_engine import DataEngine
@pytest.mark.parametrize("num_samples", [16])
def test_map_dataset(num_samples: int):
data_args = DataArguments(dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args)
data_args = DataArguments(train_dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args.train_dataset)
original_data = load_dataset("llamafactory/v1-sft-demo", split="train")
indexes = random.choices(range(len(data_engine)), k=num_samples)
for index in indexes:

4
tests_v1/core/utils/test_batching.py
View File

@@ -19,8 +19,8 @@ from llamafactory.v1.core.utils.batching import BatchGenerator
def test_normal_batching():
data_args = DataArguments(dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args=data_args)
data_args = DataArguments(train_dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args.train_dataset)
model_args = ModelArguments(model="llamafactory/tiny-random-qwen3")
model_engine = ModelEngine(model_args=model_args)
training_args = TrainingArguments(

10
tests_v1/core/utils/test_rendering.py
View File

@@ -111,8 +111,8 @@ def test_chatml_parse():
def test_chatml_rendering_remote(num_samples: int):
tokenizer: Processor = AutoTokenizer.from_pretrained("llamafactory/tiny-random-qwen3")
renderer = Renderer(template="chatml", processor=tokenizer)
data_args = DataArguments(dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args)
data_args = DataArguments(train_dataset="llamafactory/v1-sft-demo")
data_engine = DataEngine(data_args.train_dataset)
for index in range(num_samples):
v1_inputs = renderer.render_messages(data_engine[index]["messages"], is_generate=True)
prefix = tokenizer.encode("<|im_start|>user\n", add_special_tokens=False)
@@ -167,8 +167,8 @@ def test_qwen3_nothink_parse():
def test_qwen3_nothink_rendering_remote(num_samples: int):
tokenizer: Processor = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
renderer = Renderer(template="qwen3_nothink", processor=tokenizer)
data_args = DataArguments(dataset="llamafactory/reason-tool-use-demo-1500")
data_engine = DataEngine(data_args)
data_args = DataArguments(train_dataset="llamafactory/reason-tool-use-demo-1500")
data_engine = DataEngine(data_args.train_dataset)
for index in range(num_samples):
v1_inputs = renderer.render_messages(data_engine[index]["messages"], tools=data_engine[index]["tools"])
prefix_text = (
@@ -213,7 +213,7 @@ def test_process_dpo_samples():
model_inputs = renderer.process_samples(samples)
assert len(model_inputs) == 1
assert model_inputs[0]["input_ids"] == hf_inputs * 2
assert model_inputs[0]["token_type_ids"] == [0] * len(hf_inputs) + [1] * len(hf_inputs)
assert model_inputs[0]["token_type_ids"] == [1] * len(hf_inputs) + [2] * len(hf_inputs)
assert model_inputs[0]["extra_info"] == "test"
assert model_inputs[0]["_dataset_name"] == "default"