[data] fix loader (#7207)

* fix dataloader

* add test case

* fix type

* fix ci

* fix ci

* fix ci

* disable overwrite cache in ci

Former-commit-id: e84af0e140b1aafd1a6d6fe185a8e41c8fc5f831

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)