fix up

Former-commit-id: 43a56cb331fae899ca35b0c312730d4ab79d0c42

src/llamafactory/train/dpo/workflow.py

@@ -17,7 +17,7 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.ploting import plot_loss
 from ...hparams import ModelArguments
@@ -70,8 +70,8 @@ def run_dpo(
         finetuning_args=finetuning_args,
         data_collator=data_collator,
         callbacks=callbacks,
-        **tokenizer_module,
         **dataset_module,
+        **tokenizer_module,
     )
 
     # Training

src/llamafactory/train/kto/workflow.py

@@ -17,7 +17,7 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import KTODataCollatorWithPadding, get_dataset, split_dataset
+from ...data import KTODataCollatorWithPadding, get_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.ploting import plot_loss
 from ...hparams import ModelArguments
@@ -67,8 +67,8 @@ def run_kto(
         finetuning_args=finetuning_args,
         data_collator=data_collator,
         callbacks=callbacks,
-        **tokenizer_module,
         **dataset_module,
+        **tokenizer_module,
     )
 
     # Training

src/llamafactory/train/ppo/trainer.py

@@ -77,9 +77,13 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
         ref_model: Optional["AutoModelForCausalLMWithValueHead"],
         tokenizer: "PreTrainedTokenizer",
         processor: Optional["ProcessorMixin"],
-        dataset: "Dataset",
         data_collator: "DataCollatorWithPadding",
+        train_dataset: Optional["Dataset"] = None,
+        eval_dataset: Optional["Dataset"] = None,
     ) -> None:
+        if eval_dataset is not None:
+            raise NotImplementedError("PPOTrainer does not support eval dataset yet.")
+
         backward_batch_size = training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps
         ppo_config = PPOConfig(
             model_name=model_args.model_name_or_path,
@@ -115,7 +119,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
             num_training_steps = training_args.max_steps
         else:
             total_train_batch_size = backward_batch_size * finetuning_args.ppo_buffer_size * training_args.world_size
-            num_training_steps = training_args.num_train_epochs * math.ceil(len(dataset) / total_train_batch_size)
+            num_training_steps = training_args.num_train_epochs * math.ceil(len(train_dataset) / total_train_batch_size)
 
         optimizer = self.create_optimizer(model, training_args, finetuning_args)
         scheduler = self.create_scheduler(training_args, num_training_steps, optimizer)
@@ -126,7 +130,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
             model=model,
             ref_model=ref_model,
             tokenizer=tokenizer,
-            dataset=dataset,
+            dataset=train_dataset,
             data_collator=data_collator,
             lr_scheduler=scheduler,
         )

src/llamafactory/train/ppo/workflow.py

@@ -63,8 +63,8 @@ def run_ppo(
         model=model,
         reward_model=reward_model,
         ref_model=ref_model,
-        dataset=dataset_module["train_dataset"],
         data_collator=data_collator,
+        **dataset_module,
         **tokenizer_module,
     )
 

src/llamafactory/train/pt/workflow.py

@@ -20,7 +20,7 @@ from typing import TYPE_CHECKING, List, Optional
 
 from transformers import DataCollatorForLanguageModeling
 
-from ...data import get_dataset, split_dataset
+from ...data import get_dataset
 from ...extras.ploting import plot_loss
 from ...model import load_model, load_tokenizer
 from ..trainer_utils import create_modelcard_and_push
@@ -53,8 +53,8 @@ def run_pt(
         finetuning_args=finetuning_args,
         data_collator=data_collator,
         callbacks=callbacks,
-        **tokenizer_module,
         **dataset_module,
+        **tokenizer_module,
     )
 
     # Training

src/llamafactory/train/rm/workflow.py

@@ -17,7 +17,7 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
+from ...data import PairwiseDataCollatorWithPadding, get_dataset
 from ...extras.ploting import plot_loss
 from ...model import load_model, load_tokenizer
 from ..callbacks import fix_valuehead_checkpoint
@@ -56,8 +56,8 @@ def run_rm(
         data_collator=data_collator,
         callbacks=callbacks,
         compute_metrics=compute_accuracy,
-        **tokenizer_module,
         **dataset_module,
+        **tokenizer_module,
     )
 
     # Training

src/llamafactory/train/sft/workflow.py

@@ -17,7 +17,7 @@
 
 from typing import TYPE_CHECKING, List, Optional
 
-from ...data import SFTDataCollatorWith4DAttentionMask, get_dataset, split_dataset
+from ...data import SFTDataCollatorWith4DAttentionMask, get_dataset
 from ...extras.constants import IGNORE_INDEX
 from ...extras.misc import get_logits_processor
 from ...extras.ploting import plot_loss
@@ -75,8 +75,8 @@ def run_sft(
         callbacks=callbacks,
         compute_metrics=ComputeMetrics(tokenizer) if training_args.predict_with_generate else compute_accuracy,
         preprocess_logits_for_metrics=None if training_args.predict_with_generate else eval_logit_processor,
-        **tokenizer_module,
         **dataset_module,
+        **tokenizer_module,
     )
 
     # Keyword arguments for `model.generate`

src/llamafactory/train/trainer_utils.py

@@ -79,7 +79,7 @@ def create_modelcard_and_push(
         "tags": ["llama-factory", finetuning_args.finetuning_type],
     }
     if data_args.dataset is not None:
-        kwargs["dataset"] = [dataset.strip() for dataset in data_args.dataset.split(",")]
+        kwargs["dataset"] = data_args.dataset
 
     if model_args.use_unsloth:
         kwargs["tags"] = kwargs["tags"] + ["unsloth"]