423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-17 04:10:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

[misc] fix packing and eval plot (#7623)

Browse Source
This commit is contained in:
hoshi-hiyouga
2025-04-07 18:20:57 +08:00
committed by GitHub
parent 5115dc8c7f
commit c3c0efbaa0
70 changed files with 288 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/data/processor/test_feedback.py
View File

@@ -25,10 +25,10 @@ from llamafactory.train.test_utils import load_dataset_module
DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"model_name_or_path": TINY_LLAMA3,
"stage": "kto",
"do_train": True,
"finetuning_type": "full",
@@ -45,7 +45,7 @@ TRAIN_ARGS = {
@pytest.mark.parametrize("num_samples", [16])
def test_feedback_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="kto_en_demo", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:

6
tests/data/processor/test_pairwise.py
View File

@@ -25,10 +25,10 @@ from llamafactory.train.test_utils import load_dataset_module
DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"model_name_or_path": TINY_LLAMA3,
"stage": "rm",
"do_train": True,
"finetuning_type": "full",
@@ -54,7 +54,7 @@ def _convert_sharegpt_to_openai(messages: list[dict[str, str]]) -> list[dict[str
@pytest.mark.parametrize("num_samples", [16])
def test_pairwise_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="dpo_en_demo", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:

12
tests/data/processor/test_supervised.py
View File

@@ -25,12 +25,12 @@ from llamafactory.train.test_utils import load_dataset_module
DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"model_name_or_path": TINY_LLAMA3,
"stage": "sft",
"do_train": True,
"finetuning_type": "full",
@@ -45,7 +45,7 @@ TRAIN_ARGS = {
@pytest.mark.parametrize("num_samples", [16])
def test_supervised_single_turn(num_samples: int):
train_dataset = load_dataset_module(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(TINY_DATA, split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
@@ -66,7 +66,7 @@ def test_supervised_multi_turn(num_samples: int):
train_dataset = load_dataset_module(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)[
"train_dataset"
]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
@@ -79,7 +79,7 @@ def test_supervised_train_on_prompt(num_samples: int):
train_dataset = load_dataset_module(
dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", train_on_prompt=True, **TRAIN_ARGS
)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
@@ -93,7 +93,7 @@ def test_supervised_mask_history(num_samples: int):
train_dataset = load_dataset_module(
dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", mask_history=True, **TRAIN_ARGS
)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:

6
tests/data/processor/test_unsupervised.py
View File

@@ -24,12 +24,12 @@ from llamafactory.train.test_utils import load_dataset_module
DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"model_name_or_path": TINY_LLAMA3,
"stage": "ppo",
"do_train": True,
"finetuning_type": "full",
@@ -48,7 +48,7 @@ TRAIN_ARGS = {
@pytest.mark.parametrize("num_samples", [16])
def test_unsupervised_data(num_samples: int):
train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes: