update to v0.9.2

src/llamafactory/train/test_utils.py

@@ -26,10 +26,11 @@ from ..model import load_model, load_tokenizer
 
 
 if TYPE_CHECKING:
-    from datasets import Dataset
     from peft import LoraModel
     from transformers import PreTrainedModel
 
+    from ..data.data_utils import DatasetModule
+
 
 def compare_model(model_a: "torch.nn.Module", model_b: "torch.nn.Module", diff_keys: Sequence[str] = []) -> None:
     state_dict_a = model_a.state_dict()
@@ -101,12 +102,12 @@ def load_reference_model(
     return model
 
 
-def load_train_dataset(**kwargs) -> "Dataset":
+def load_dataset_module(**kwargs) -> "DatasetModule":
     model_args, data_args, training_args, _, _ = get_train_args(kwargs)
     tokenizer_module = load_tokenizer(model_args)
     template = get_template_and_fix_tokenizer(tokenizer_module["tokenizer"], data_args)
     dataset_module = get_dataset(template, model_args, data_args, training_args, kwargs["stage"], **tokenizer_module)
-    return dataset_module["train_dataset"]
+    return dataset_module
 
 
 def patch_valuehead_model() -> None:

src/llamafactory/train/trainer_utils.py

@@ -617,6 +617,7 @@ def get_swanlab_callback(finetuning_args: "FinetuningArguments") -> "TrainerCall
         experiment_name=finetuning_args.swanlab_run_name,
         mode=finetuning_args.swanlab_mode,
         config={"Framework": "🦙LlamaFactory"},
+        logdir=finetuning_args.swanlab_logdir,
     )
     return swanlab_callback
 

src/llamafactory/train/tuner.py

@@ -86,6 +86,9 @@ def _training_function(config: Dict[str, Any]) -> None:
 
 def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["TrainerCallback"]] = None) -> None:
     args = read_args(args)
+    if "-h" in args or "--help" in args:
+        get_train_args(args)
+
     ray_args = get_ray_args(args)
     callbacks = callbacks or []
     if ray_args.use_ray:
@@ -182,6 +185,7 @@ def export_model(args: Optional[Dict[str, Any]] = None) -> None:
     except Exception as e:
         logger.warning_rank0(f"Cannot save tokenizer, please copy the files manually: {e}.")
 
-    with open(os.path.join(model_args.export_dir, "Modelfile"), "w", encoding="utf-8") as f:
+    ollama_modelfile = os.path.join(model_args.export_dir, "Modelfile")
+    with open(ollama_modelfile, "w", encoding="utf-8") as f:
         f.write(template.get_ollama_modelfile(tokenizer))
-        logger.info_rank0(f"Saved ollama modelfile to {model_args.export_dir}.")
+        logger.info_rank0(f"Ollama modelfile saved in {ollama_modelfile}")

src/llamafactory/webui/locales.py

@@ -1894,6 +1894,28 @@ LOCALES = {
             "info": "クラウド版またはオフライン版 SwanLab を使用します。",
         },
     },
+    "swanlab_logdir": {
+        "en": {
+            "label": "SwanLab log directory",
+            "info": "The log directory for SwanLab.",
+        },
+        "ru": {
+            "label": "SwanLab 로그 디렉토리",
+            "info": "SwanLab의 로그 디렉토리.",
+        },
+        "zh": {
+            "label": "SwanLab 日志目录",
+            "info": "SwanLab 的日志目录。",
+        },
+        "ko": {
+            "label": "SwanLab 로그 디렉토리",
+            "info": "SwanLab의 로그 디렉토리.",
+        },
+        "ja": {
+            "label": "SwanLab ログ ディレクトリ",
+            "info": "SwanLab のログ ディレクトリ。",
+        },
+    },
     "cmd_preview_btn": {
         "en": {
             "value": "Preview command",

tests/data/processor/test_feedback.py

@@ -20,7 +20,7 @@ from datasets import load_dataset
 from transformers import AutoTokenizer
 
 from llamafactory.extras.constants import IGNORE_INDEX
-from llamafactory.train.test_utils import load_train_dataset
+from llamafactory.train.test_utils import load_dataset_module
 
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
@@ -36,7 +36,6 @@ TRAIN_ARGS = {
     "dataset_dir": "REMOTE:" + DEMO_DATA,
     "template": "llama3",
     "cutoff_len": 8192,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,
@@ -45,7 +44,7 @@ TRAIN_ARGS = {
 
 @pytest.mark.parametrize("num_samples", [16])
 def test_feedback_data(num_samples: int):
-    train_dataset = load_train_dataset(**TRAIN_ARGS)
+    train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
     original_data = load_dataset(DEMO_DATA, name="kto_en_demo", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)

tests/data/processor/test_pairwise.py

@@ -21,7 +21,7 @@ from datasets import load_dataset
 from transformers import AutoTokenizer
 
 from llamafactory.extras.constants import IGNORE_INDEX
-from llamafactory.train.test_utils import load_train_dataset
+from llamafactory.train.test_utils import load_dataset_module
 
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
@@ -37,7 +37,6 @@ TRAIN_ARGS = {
     "dataset_dir": "REMOTE:" + DEMO_DATA,
     "template": "llama3",
     "cutoff_len": 8192,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,
@@ -55,7 +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_train_dataset(**TRAIN_ARGS)
+    train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
     original_data = load_dataset(DEMO_DATA, name="dpo_en_demo", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)

tests/data/processor/test_supervised.py

@@ -20,7 +20,7 @@ from datasets import load_dataset
 from transformers import AutoTokenizer
 
 from llamafactory.extras.constants import IGNORE_INDEX
-from llamafactory.train.test_utils import load_train_dataset
+from llamafactory.train.test_utils import load_dataset_module
 
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
@@ -36,7 +36,6 @@ TRAIN_ARGS = {
     "finetuning_type": "full",
     "template": "llama3",
     "cutoff_len": 8192,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,
@@ -45,7 +44,7 @@ TRAIN_ARGS = {
 
 @pytest.mark.parametrize("num_samples", [16])
 def test_supervised_single_turn(num_samples: int):
-    train_dataset = load_train_dataset(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)
+    train_dataset = load_dataset_module(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)["train_dataset"]
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
     original_data = load_dataset(TINY_DATA, split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
@@ -64,7 +63,9 @@ def test_supervised_single_turn(num_samples: int):
 
 @pytest.mark.parametrize("num_samples", [8])
 def test_supervised_multi_turn(num_samples: int):
-    train_dataset = load_train_dataset(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)
+    train_dataset = load_dataset_module(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)[
+        "train_dataset"
+    ]
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
@@ -75,9 +76,9 @@ def test_supervised_multi_turn(num_samples: int):
 
 @pytest.mark.parametrize("num_samples", [4])
 def test_supervised_train_on_prompt(num_samples: int):
-    train_dataset = load_train_dataset(
+    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)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
@@ -89,9 +90,9 @@ def test_supervised_train_on_prompt(num_samples: int):
 
 @pytest.mark.parametrize("num_samples", [4])
 def test_supervised_mask_history(num_samples: int):
-    train_dataset = load_train_dataset(
+    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)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)

tests/data/processor/test_unsupervised.py

@@ -19,7 +19,7 @@ import pytest
 from datasets import load_dataset
 from transformers import AutoTokenizer
 
-from llamafactory.train.test_utils import load_train_dataset
+from llamafactory.train.test_utils import load_dataset_module
 
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
@@ -39,7 +39,6 @@ TRAIN_ARGS = {
     "dataset_dir": "REMOTE:" + DEMO_DATA,
     "template": "llama3",
     "cutoff_len": 8192,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,
@@ -48,7 +47,7 @@ TRAIN_ARGS = {
 
 @pytest.mark.parametrize("num_samples", [16])
 def test_unsupervised_data(num_samples: int):
-    train_dataset = load_train_dataset(**TRAIN_ARGS)
+    train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)

tests/data/test_converter.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from llamafactory.data import Role
 from llamafactory.data.converter import get_dataset_converter
 from llamafactory.data.parser import DatasetAttr

tests/data/test_loader.py

+# Copyright 2025 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+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_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
+
+TRAIN_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "stage": "sft",
+    "do_train": True,
+    "finetuning_type": "full",
+    "template": "llama3",
+    "dataset": TINY_DATA,
+    "dataset_dir": "ONLINE",
+    "cutoff_len": 8192,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+def test_load_train_only():
+    dataset_module = load_dataset_module(**TRAIN_ARGS)
+    assert dataset_module.get("train_dataset") is not None
+    assert dataset_module.get("eval_dataset") is None
+
+
+def test_load_val_size():
+    dataset_module = load_dataset_module(val_size=0.1, **TRAIN_ARGS)
+    assert dataset_module.get("train_dataset") is not None
+    assert dataset_module.get("eval_dataset") is not None
+
+
+def test_load_eval_data():
+    dataset_module = load_dataset_module(eval_dataset=TINY_DATA, **TRAIN_ARGS)
+    assert dataset_module.get("train_dataset") is not None
+    assert dataset_module.get("eval_dataset") is not None

tests/e2e/test_train.py

@@ -32,7 +32,6 @@ TRAIN_ARGS = {
     "dataset_dir": "REMOTE:" + DEMO_DATA,
     "template": "llama3",
     "cutoff_len": 1,
-    "overwrite_cache": False,
     "overwrite_output_dir": True,
     "per_device_train_batch_size": 1,
     "max_steps": 1,

tests/model/model_utils/test_checkpointing.py

@@ -33,7 +33,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,

tests/model/model_utils/test_visual.py

@@ -20,23 +20,16 @@ from llamafactory.hparams import FinetuningArguments, ModelArguments
 from llamafactory.model.adapter import init_adapter
 
 
-@pytest.mark.parametrize(
-    "freeze_vision_tower,freeze_multi_modal_projector,train_mm_proj_only",
-    [
-        (False, False, False),
-        (False, True, False),
-        (True, False, False),
-        (True, True, False),
-        (True, False, True),
-    ],
-)
-def test_visual_full(freeze_vision_tower: bool, freeze_multi_modal_projector: bool, train_mm_proj_only: bool):
+@pytest.mark.parametrize("freeze_vision_tower", (False, True))
+@pytest.mark.parametrize("freeze_multi_modal_projector", (False, True))
+@pytest.mark.parametrize("freeze_language_model", (False, True))
+def test_visual_full(freeze_vision_tower: bool, freeze_multi_modal_projector: bool, freeze_language_model: bool):
     model_args = ModelArguments(model_name_or_path="Qwen/Qwen2-VL-2B-Instruct")
     finetuning_args = FinetuningArguments(
         finetuning_type="full",
         freeze_vision_tower=freeze_vision_tower,
         freeze_multi_modal_projector=freeze_multi_modal_projector,
-        train_mm_proj_only=train_mm_proj_only,
+        freeze_language_model=freeze_language_model,
     )
     config = AutoConfig.from_pretrained(model_args.model_name_or_path)
     with torch.device("meta"):
@@ -49,10 +42,10 @@ def test_visual_full(freeze_vision_tower: bool, freeze_multi_modal_projector: bo
         elif "visual.merger" in name:
             assert param.requires_grad != freeze_multi_modal_projector
         else:
-            assert param.requires_grad != train_mm_proj_only
+            assert param.requires_grad != freeze_language_model
 
 
-@pytest.mark.parametrize("freeze_vision_tower", [False, True])
+@pytest.mark.parametrize("freeze_vision_tower", (False, True))
 def test_visual_lora(freeze_vision_tower: bool):
     model_args = ModelArguments(model_name_or_path="Qwen/Qwen2-VL-2B-Instruct")
     finetuning_args = FinetuningArguments(finetuning_type="lora", freeze_vision_tower=freeze_vision_tower)

tests/model/test_freeze.py

@@ -30,7 +30,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,

tests/model/test_full.py

@@ -30,7 +30,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,

tests/model/test_lora.py

@@ -42,7 +42,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,

tests/model/test_pissa.py

@@ -34,7 +34,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": True,
     "output_dir": "dummy_dir",
     "overwrite_output_dir": True,
     "fp16": True,

tests/train/test_sft_trainer.py

@@ -38,7 +38,6 @@ TRAIN_ARGS = {
     "dataset_dir": "ONLINE",
     "template": "llama3",
     "cutoff_len": 1024,
-    "overwrite_cache": False,
     "overwrite_output_dir": True,
     "per_device_train_batch_size": 1,
     "max_steps": 1,