diff --git a/data/README.md b/data/README.md
index 5ceae666..0f14bef8 100644
--- a/data/README.md
+++ b/data/README.md
@@ -12,7 +12,8 @@ Currently we support datasets in **alpaca** and **sharegpt** format.
   "ranking": "whether the dataset is a preference dataset or not. (default: False)",
   "subset": "the name of the subset. (optional, default: None)",
   "folder": "the name of the folder of the dataset repository on the Hugging Face hub. (optional, default: None)",
-  "num_samples": "the number of samples in the dataset used for training. (optional, default: None)",
+  "num_samples": "the number of samples in the dataset used for training. (optional, default: None)", 
+  "split": "which dataset split to use for training and evaluation (optional, default: train)",
   "columns (optional)": {
     "prompt": "the column name in the dataset containing the prompts. (default: instruction)",
     "query": "the column name in the dataset containing the queries. (default: input)",
diff --git a/data/README_zh.md b/data/README_zh.md
index 1795f352..7bf4fdba 100644
--- a/data/README_zh.md
+++ b/data/README_zh.md
@@ -13,6 +13,7 @@
   "subset": "数据集子集的名称（可选，默认：None）",
   "folder": "Hugging Face 仓库的文件夹名称（可选，默认：None）",
   "num_samples": "该数据集中用于训练的样本数量。（可选，默认：None）",
+  "split": "数据集中的要使用的训练测试集切分（可选，默认：train）",
   "columns（可选）": {
     "prompt": "数据集代表提示词的表头名称（默认：instruction）",
     "query": "数据集代表请求的表头名称（默认：input）",
diff --git a/data/dataset_info.json b/data/dataset_info.json
index f8ffd407..e4b5a384 100644
--- a/data/dataset_info.json
+++ b/data/dataset_info.json
@@ -172,9 +172,19 @@
   "deepctrl": {
     "ms_hub_url": "deepctrl/deepctrl-sft-data"
   },
-  "adgen": {
+  "adgen_train": {
     "hf_hub_url": "HasturOfficial/adgen",
     "ms_hub_url": "AI-ModelScope/adgen",
+    "split": "train",
+    "columns": {
+      "prompt": "content",
+      "response": "summary"
+    }
+  },
+  "adgen_val": {
+    "hf_hub_url": "HasturOfficial/adgen",
+    "ms_hub_url": "AI-ModelScope/adgen",
+    "split": "validation",
     "columns": {
       "prompt": "content",
       "response": "summary"
diff --git a/scripts/cal_lr.py b/scripts/cal_lr.py
index a103e082..a38f34e1 100644
--- a/scripts/cal_lr.py
+++ b/scripts/cal_lr.py
@@ -65,7 +65,7 @@ def calculate_lr(
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
     elif stage == "sft":
@@ -73,7 +73,7 @@ def calculate_lr(
     else:
         raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
-    dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
+    dataloader = DataLoader(dataset_module["eval_dataset"], batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     valid_tokens, total_tokens = 0, 0
     for batch in tqdm(dataloader):
         valid_tokens += torch.sum(batch["labels"] != IGNORE_INDEX).item()
diff --git a/scripts/cal_ppl.py b/scripts/cal_ppl.py
index 61f76922..3daa35ae 100644
--- a/scripts/cal_ppl.py
+++ b/scripts/cal_ppl.py
@@ -87,7 +87,7 @@ def cal_ppl(
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    trainset = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage, **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, is_trainable=False)
     if stage == "pt":
         data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
@@ -100,7 +100,7 @@ def cal_ppl(
     else:
         raise NotImplementedError("Stage does not supported: {}.".format(stage))
 
-    dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
+    dataloader = DataLoader(dataset_module["eval_dataset"], batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     criterion = torch.nn.CrossEntropyLoss(reduction="none")
     total_ppl = 0
     perplexities = []
diff --git a/scripts/length_cdf.py b/scripts/length_cdf.py
index 4cdf01e6..cef46416 100644
--- a/scripts/length_cdf.py
+++ b/scripts/length_cdf.py
@@ -47,10 +47,10 @@ def length_cdf(
         )
     )
     tokenizer_module = load_tokenizer(model_args)
-    trainset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
-    total_num = len(trainset)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    total_num = len(dataset_module["eval_dataset"])
     length_dict = defaultdict(int)
-    for sample in tqdm(trainset["input_ids"]):
+    for sample in tqdm(dataset_module["eval_dataset"]["input_ids"]):
         length_dict[len(sample) // interval * interval] += 1
 
     length_tuples = list(length_dict.items())
diff --git a/src/llamafactory/data/loader.py b/src/llamafactory/data/loader.py
index 8e7062db..d527d7d2 100644
--- a/src/llamafactory/data/loader.py
+++ b/src/llamafactory/data/loader.py
@@ -15,7 +15,7 @@
 import inspect
 import os
 import sys
-from typing import TYPE_CHECKING, Literal, Optional, Union
+from typing import TYPE_CHECKING, Literal, Optional, Union, Dict
 
 import numpy as np
 from datasets import load_dataset, load_from_disk
@@ -24,10 +24,10 @@ from ..extras.constants import FILEEXT2TYPE
 from ..extras.logging import get_logger
 from ..extras.misc import has_tokenized_data
 from .aligner import align_dataset
-from .data_utils import merge_dataset
+from .data_utils import merge_dataset, split_dataset
 from .parser import get_dataset_list
 from .preprocess import get_preprocess_and_print_func
-from .template import get_template_and_fix_tokenizer
+from .template import get_template_and_fix_tokenizer, Template
 
 
 if TYPE_CHECKING:
@@ -91,7 +91,7 @@ def load_single_dataset(
                 subset_name=data_name,
                 data_dir=data_dir,
                 data_files=data_files,
-                split=data_args.split,
+                split=dataset_attr.split,
                 cache_dir=cache_dir,
                 token=model_args.ms_hub_token,
                 use_streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
@@ -111,7 +111,7 @@ def load_single_dataset(
             name=data_name,
             data_dir=data_dir,
             data_files=data_files,
-            split=data_args.split,
+            split=dataset_attr.split,
             cache_dir=model_args.cache_dir,
             token=model_args.hf_hub_token,
             streaming=(data_args.streaming and (dataset_attr.load_from != "file")),
@@ -140,20 +140,17 @@ def load_single_dataset(
     return align_dataset(dataset, dataset_attr, data_args, training_args)
 
 
-def get_dataset(
+def load_and_preprocess(
     model_args: "ModelArguments",
     data_args: "DataArguments",
     training_args: "Seq2SeqTrainingArguments",
     stage: Literal["pt", "sft", "rm", "ppo", "kto"],
     tokenizer: "PreTrainedTokenizer",
+    template: "Template",
     processor: Optional["ProcessorMixin"] = None,
+    is_eval: bool = False
 ) -> Union["Dataset", "IterableDataset"]:
-    template = get_template_and_fix_tokenizer(tokenizer, data_args.template, data_args.tool_format)
-    if data_args.train_on_prompt and template.efficient_eos:
-        raise ValueError("Current template does not support `train_on_prompt`.")
-
-    # Load tokenized dataset
-    if data_args.tokenized_path is not None:
+    if not is_eval and data_args.tokenized_path is not None:
         if has_tokenized_data(data_args.tokenized_path):
             logger.warning("Loading dataset from disk will ignore other data arguments.")
             dataset = load_from_disk(data_args.tokenized_path)
@@ -165,9 +162,21 @@ def get_dataset(
         if data_args.streaming:
             raise ValueError("Turn off `streaming` when saving dataset to disk.")
 
+    if is_eval and data_args.eval_tokenized_path is not None:
+        if has_tokenized_data(data_args.eval_tokenized_path):
+            logger.warning("Loading dataset from disk will ignore other data arguments.")
+            dataset = load_from_disk(data_args.eval_tokenized_path)
+            logger.info("Loaded tokenized dataset from {}.".format(data_args.eval_tokenized_path))
+            if data_args.streaming:
+                dataset = dataset.to_iterable_dataset()
+            return dataset
+
+        if data_args.streaming:
+            raise ValueError("Turn off `streaming` when saving dataset to disk.")
+
     with training_args.main_process_first(desc="load dataset"):
         all_datasets = []
-        for dataset_attr in get_dataset_list(data_args):
+        for dataset_attr in get_dataset_list(data_args, data_args.eval_dataset if is_eval else data_args.dataset):
             if (stage == "rm" and dataset_attr.ranking is False) or (stage != "rm" and dataset_attr.ranking is True):
                 raise ValueError("The dataset is not applicable in the current training stage.")
 
@@ -190,13 +199,20 @@ def get_dataset(
 
         dataset = dataset.map(preprocess_func, batched=True, remove_columns=column_names, **kwargs)
 
-        if data_args.tokenized_path is not None:
+        if not is_eval and data_args.tokenized_path is not None:
             if training_args.should_save:
                 dataset.save_to_disk(data_args.tokenized_path)
                 logger.info("Tokenized dataset saved at {}.".format(data_args.tokenized_path))
                 logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.tokenized_path))
 
             sys.exit(0)
+        if is_eval and data_args.eval_tokenized_path is not None:
+            if training_args.should_save:
+                dataset.save_to_disk(data_args.eval_tokenized_path)
+                logger.info("Tokenized dataset saved at {}.".format(data_args.eval_tokenized_path))
+                logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.eval_tokenized_path))
+
+            sys.exit(0)
 
         if training_args.should_log:
             try:
@@ -208,3 +224,24 @@ def get_dataset(
                     raise RuntimeError("Cannot find valid samples, check `data/README.md` for the data format.")
 
         return dataset
+
+
+def get_dataset(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    stage: Literal["pt", "sft", "rm", "ppo", "kto"],
+    tokenizer: "PreTrainedTokenizer",
+    processor: Optional["ProcessorMixin"] = None
+) -> Dict[str, "Dataset"]:
+    template = get_template_and_fix_tokenizer(tokenizer, data_args.template, data_args.tool_format)
+    if data_args.train_on_prompt and template.efficient_eos:
+        raise ValueError("Current template does not support `train_on_prompt`.")
+
+    train_dataset = load_and_preprocess(model_args, data_args, training_args, stage, tokenizer, template, processor)
+
+    if data_args.eval_dataset or data_args.eval_tokenized_path:
+        eval_dataset = load_and_preprocess(model_args, data_args, training_args, stage, tokenizer, template, processor, True)
+        return {"train_dataset": train_dataset, "eval_dataset": eval_dataset}
+    else:
+        return split_dataset(train_dataset, data_args, training_args)
diff --git a/src/llamafactory/data/parser.py b/src/llamafactory/data/parser.py
index 5ae79774..c810ec8b 100644
--- a/src/llamafactory/data/parser.py
+++ b/src/llamafactory/data/parser.py
@@ -40,6 +40,7 @@ class DatasetAttr:
     subset: Optional[str] = None
     folder: Optional[str] = None
     num_samples: Optional[int] = None
+    split: Optional[str] = "train"
     # common columns
     system: Optional[str] = None
     tools: Optional[str] = None
@@ -71,9 +72,9 @@ class DatasetAttr:
         setattr(self, key, obj.get(key, default))
 
 
-def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
-    if data_args.dataset is not None:
-        dataset_names = [ds.strip() for ds in data_args.dataset.split(",")]
+def get_dataset_list(data_args: "DataArguments", dataset: "str" = None) -> List["DatasetAttr"]:
+    if dataset is not None:
+        dataset_names = [ds.strip() for ds in dataset.split(",")]
     else:
         dataset_names = []
 
@@ -122,6 +123,8 @@ def get_dataset_list(data_args: "DataArguments") -> List["DatasetAttr"]:
         dataset_attr.set_attr("subset", dataset_info[name])
         dataset_attr.set_attr("folder", dataset_info[name])
         dataset_attr.set_attr("num_samples", dataset_info[name])
+        if "split" in dataset_info[name]:
+            dataset_attr.set_attr("split", dataset_info[name])
 
         if "columns" in dataset_info[name]:
             column_names = ["system", "tools", "images", "chosen", "rejected", "kto_tag"]
diff --git a/src/llamafactory/hparams/data_args.py b/src/llamafactory/hparams/data_args.py
index a1025af7..7f7e62cd 100644
--- a/src/llamafactory/hparams/data_args.py
+++ b/src/llamafactory/hparams/data_args.py
@@ -33,6 +33,11 @@ class DataArguments:
         default=None,
         metadata={"help": "The name of provided dataset(s) to use. Use commas to separate multiple datasets."},
     )
+    eval_dataset: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of provided dataset(s) to use for eval during training. "
+                          "Use commas to separate multiple datasets."},
+    )
     dataset_dir: str = field(
         default="data",
         metadata={"help": "Path to the folder containing the datasets."},
@@ -105,6 +110,10 @@ class DataArguments:
         default=None,
         metadata={"help": "Path to save or load the tokenized datasets."},
     )
+    eval_tokenized_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to save or load the tokenized eval datasets."},
+    )
 
     def __post_init__(self):
         if self.streaming and self.val_size > 1e-6 and self.val_size < 1:
diff --git a/src/llamafactory/train/dpo/workflow.py b/src/llamafactory/train/dpo/workflow.py
index 431b5285..c004363a 100644
--- a/src/llamafactory/train/dpo/workflow.py
+++ b/src/llamafactory/train/dpo/workflow.py
@@ -41,7 +41,7 @@ def run_dpo(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="rm", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
     data_collator = PairwiseDataCollatorWithPadding(
@@ -71,7 +71,7 @@ def run_dpo(
         data_collator=data_collator,
         callbacks=callbacks,
         **tokenizer_module,
-        **split_dataset(dataset, data_args, training_args),
+        **dataset_module,
     )
 
     # Training
diff --git a/src/llamafactory/train/kto/workflow.py b/src/llamafactory/train/kto/workflow.py
index 8182a184..b2d0c82e 100644
--- a/src/llamafactory/train/kto/workflow.py
+++ b/src/llamafactory/train/kto/workflow.py
@@ -41,7 +41,7 @@ def run_kto(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="kto", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="kto", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
     data_collator = KTODataCollatorWithPadding(
@@ -68,7 +68,7 @@ def run_kto(
         data_collator=data_collator,
         callbacks=callbacks,
         **tokenizer_module,
-        **split_dataset(dataset, data_args, training_args),
+        **dataset_module,
     )
 
     # Training
diff --git a/src/llamafactory/train/ppo/workflow.py b/src/llamafactory/train/ppo/workflow.py
index f52b80d6..53d9f18f 100644
--- a/src/llamafactory/train/ppo/workflow.py
+++ b/src/llamafactory/train/ppo/workflow.py
@@ -43,7 +43,7 @@ def run_ppo(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="ppo", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="ppo", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train, add_valuehead=True)
 
     tokenizer.padding_side = "left"  # use left-padding in generation while using right-padding in training
@@ -63,7 +63,7 @@ def run_ppo(
         model=model,
         reward_model=reward_model,
         ref_model=ref_model,
-        dataset=dataset,
+        dataset=dataset_module["train_dataset"],
         data_collator=data_collator,
         **tokenizer_module,
     )
diff --git a/src/llamafactory/train/pt/workflow.py b/src/llamafactory/train/pt/workflow.py
index b84a0e7d..2f27d6cd 100644
--- a/src/llamafactory/train/pt/workflow.py
+++ b/src/llamafactory/train/pt/workflow.py
@@ -42,7 +42,7 @@ def run_pt(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="pt", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="pt", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
     data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
 
@@ -54,7 +54,7 @@ def run_pt(
         data_collator=data_collator,
         callbacks=callbacks,
         **tokenizer_module,
-        **split_dataset(dataset, data_args, training_args),
+        **dataset_module,
     )
 
     # Training
diff --git a/src/llamafactory/train/rm/workflow.py b/src/llamafactory/train/rm/workflow.py
index 384814cc..54fa7fd0 100644
--- a/src/llamafactory/train/rm/workflow.py
+++ b/src/llamafactory/train/rm/workflow.py
@@ -41,7 +41,7 @@ def run_rm(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="rm", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train, add_valuehead=True)
     data_collator = PairwiseDataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
 
@@ -57,7 +57,7 @@ def run_rm(
         callbacks=callbacks,
         compute_metrics=compute_accuracy,
         **tokenizer_module,
-        **split_dataset(dataset, data_args, training_args),
+        **dataset_module,
     )
 
     # Training
@@ -81,7 +81,7 @@ def run_rm(
 
     # Predict
     if training_args.do_predict:
-        predict_results = trainer.predict(dataset, metric_key_prefix="predict")
+        predict_results = trainer.predict(dataset_module["eval_dataset"], metric_key_prefix="predict")
         trainer.log_metrics("predict", predict_results.metrics)
         trainer.save_metrics("predict", predict_results.metrics)
         trainer.save_predictions(predict_results)
diff --git a/src/llamafactory/train/sft/workflow.py b/src/llamafactory/train/sft/workflow.py
index dea3c1a8..b0bacc33 100644
--- a/src/llamafactory/train/sft/workflow.py
+++ b/src/llamafactory/train/sft/workflow.py
@@ -43,7 +43,7 @@ def run_sft(
 ):
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    dataset = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
 
     if training_args.predict_with_generate:
@@ -76,7 +76,7 @@ def run_sft(
         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,
-        **split_dataset(dataset, data_args, training_args),
+        **dataset_module,
     )
 
     # Keyword arguments for `model.generate`
@@ -105,12 +105,12 @@ def run_sft(
 
     # Predict
     if training_args.do_predict:
-        predict_results = trainer.predict(dataset, metric_key_prefix="predict", **gen_kwargs)
+        predict_results = trainer.predict(dataset_module["eval_dataset"], metric_key_prefix="predict", **gen_kwargs)
         if training_args.predict_with_generate:  # predict_loss will be wrong if predict_with_generate is enabled
             predict_results.metrics.pop("predict_loss", None)
         trainer.log_metrics("predict", predict_results.metrics)
         trainer.save_metrics("predict", predict_results.metrics)
-        trainer.save_predictions(dataset, predict_results)
+        trainer.save_predictions(dataset_module["eval_dataset"], predict_results)
 
     # Create model card
     create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)
diff --git a/tests/data/test_supervised.py b/tests/data/test_supervised.py
index 9cb49615..7ad52ee8 100644
--- a/tests/data/test_supervised.py
+++ b/tests/data/test_supervised.py
@@ -47,7 +47,7 @@ def test_supervised(num_samples: int):
     model_args, data_args, training_args, _, _ = get_train_args(TRAIN_ARGS)
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]
-    tokenized_data = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    dataset_module = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
 
     ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
 
@@ -63,5 +63,5 @@ def test_supervised(num_samples: int):
             {"role": "assistant", "content": original_data[index]["output"]},
         ]
         templated_result = ref_tokenizer.apply_chat_template(messages, tokenize=False)
-        decoded_result = tokenizer.decode(tokenized_data["input_ids"][index])
+        decoded_result = tokenizer.decode(dataset_module["train_dataset"]["input_ids"][index])
         assert templated_result == decoded_result