[breaking change] refactor data pipeline (#6901)

* refactor data

* rename file

Former-commit-id: 7a1a4ce6451cb782573d0bd9dd27a5e443e3a18b

tests/data/processor/test_feedback.py

@@ -0,0 +1,59 @@
+# 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
+import random
+
+import pytest
+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
+
+
+DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
+
+TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+TRAIN_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "stage": "kto",
+    "do_train": True,
+    "finetuning_type": "full",
+    "dataset": "kto_en_demo",
+    "dataset_dir": "REMOTE:" + DEMO_DATA,
+    "template": "llama3",
+    "cutoff_len": 8192,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+@pytest.mark.parametrize("num_samples", [16])
+def test_feedback_data(num_samples: int):
+    train_dataset = load_train_dataset(**TRAIN_ARGS)
+    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)
+    for index in indexes:
+        messages = original_data["messages"][index]
+        ref_input_ids = ref_tokenizer.apply_chat_template(messages)
+        prompt_len = len(ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True))
+        ref_labels = [IGNORE_INDEX] * prompt_len + ref_input_ids[prompt_len:]
+        assert train_dataset["input_ids"][index] == ref_input_ids
+        assert train_dataset["labels"][index] == ref_labels
+        assert train_dataset["kto_tags"][index] == original_data["label"][index]

tests/data/processor/test_pairwise.py

@@ -0,0 +1,78 @@
+# 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
+import random
+from typing import Dict, List
+
+import pytest
+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
+
+
+DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
+
+TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+TRAIN_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "stage": "rm",
+    "do_train": True,
+    "finetuning_type": "full",
+    "dataset": "dpo_en_demo",
+    "dataset_dir": "REMOTE:" + DEMO_DATA,
+    "template": "llama3",
+    "cutoff_len": 8192,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+def _convert_sharegpt_to_openai(messages: List[Dict[str, str]]) -> List[Dict[str, str]]:
+    role_mapping = {"human": "user", "gpt": "assistant", "system": "system"}
+    new_messages = []
+    for message in messages:
+        new_messages.append({"role": role_mapping[message["from"]], "content": message["value"]})
+
+    return new_messages
+
+
+@pytest.mark.parametrize("num_samples", [16])
+def test_pairwise_data(num_samples: int):
+    train_dataset = load_train_dataset(**TRAIN_ARGS)
+    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)
+    for index in indexes:
+        chosen_messages = original_data["conversations"][index] + [original_data["chosen"][index]]
+        rejected_messages = original_data["conversations"][index] + [original_data["rejected"][index]]
+        chosen_messages = _convert_sharegpt_to_openai(chosen_messages)
+        rejected_messages = _convert_sharegpt_to_openai(rejected_messages)
+        ref_chosen_input_ids = ref_tokenizer.apply_chat_template(chosen_messages)
+        chosen_prompt_len = len(ref_tokenizer.apply_chat_template(chosen_messages[:-1], add_generation_prompt=True))
+        ref_chosen_labels = [IGNORE_INDEX] * chosen_prompt_len + ref_chosen_input_ids[chosen_prompt_len:]
+        ref_rejected_input_ids = ref_tokenizer.apply_chat_template(rejected_messages)
+        rejected_prompt_len = len(
+            ref_tokenizer.apply_chat_template(rejected_messages[:-1], add_generation_prompt=True)
+        )
+        ref_rejected_labels = [IGNORE_INDEX] * rejected_prompt_len + ref_rejected_input_ids[rejected_prompt_len:]
+        assert train_dataset["chosen_input_ids"][index] == ref_chosen_input_ids
+        assert train_dataset["chosen_labels"][index] == ref_chosen_labels
+        assert train_dataset["rejected_input_ids"][index] == ref_rejected_input_ids
+        assert train_dataset["rejected_labels"][index] == ref_rejected_labels

tests/data/processor/test_processor_utils.py

@@ -0,0 +1,35 @@
+# 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 typing import Tuple
+
+import pytest
+
+from llamafactory.data.processor.processor_utils import infer_seqlen
+
+
+@pytest.mark.parametrize(
+    "test_input,test_output",
+    [
+        ((3000, 2000, 1000), (600, 400)),
+        ((2000, 3000, 1000), (400, 600)),
+        ((1000, 100, 1000), (900, 100)),
+        ((100, 1000, 1000), (100, 900)),
+        ((100, 500, 1000), (100, 500)),
+        ((500, 100, 1000), (500, 100)),
+        ((10, 10, 1000), (10, 10)),
+    ],
+)
+def test_infer_seqlen(test_input: Tuple[int, int, int], test_output: Tuple[int, int]):
+    assert test_output == infer_seqlen(*test_input)

tests/data/processor/test_supervised.py

@@ -0,0 +1,104 @@
+# 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
+import random
+
+import pytest
+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
+
+
+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",
+    "cutoff_len": 8192,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+@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)
+    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)
+    for index in indexes:
+        prompt = original_data["instruction"][index]
+        if original_data["input"][index]:
+            prompt += "\n" + original_data["input"][index]
+
+        messages = [
+            {"role": "user", "content": prompt},
+            {"role": "assistant", "content": original_data["output"][index]},
+        ]
+        ref_input_ids = ref_tokenizer.apply_chat_template(messages)
+        assert train_dataset["input_ids"][index] == ref_input_ids
+
+
+@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)
+    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)
+    for index in indexes:
+        ref_input_ids = ref_tokenizer.apply_chat_template(original_data["messages"][index])
+        assert train_dataset["input_ids"][index] == ref_input_ids
+
+
+@pytest.mark.parametrize("num_samples", [4])
+def test_supervised_train_on_prompt(num_samples: int):
+    train_dataset = load_train_dataset(
+        dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", train_on_prompt=True, **TRAIN_ARGS
+    )
+    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)
+    for index in indexes:
+        ref_ids = ref_tokenizer.apply_chat_template(original_data["messages"][index])
+        assert train_dataset["input_ids"][index] == ref_ids
+        assert train_dataset["labels"][index] == ref_ids
+
+
+@pytest.mark.parametrize("num_samples", [4])
+def test_supervised_mask_history(num_samples: int):
+    train_dataset = load_train_dataset(
+        dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", mask_history=True, **TRAIN_ARGS
+    )
+    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)
+    for index in indexes:
+        messages = original_data["messages"][index]
+        ref_input_ids = ref_tokenizer.apply_chat_template(messages)
+        prompt_len = len(ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True))
+        ref_label_ids = [IGNORE_INDEX] * prompt_len + ref_input_ids[prompt_len:]
+        assert train_dataset["input_ids"][index] == ref_input_ids
+        assert train_dataset["labels"][index] == ref_label_ids

tests/data/processor/test_unsupervised.py

@@ -0,0 +1,61 @@
+# 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
+import random
+
+import pytest
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+from llamafactory.train.test_utils import load_train_dataset
+
+
+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": "ppo",
+    "do_train": True,
+    "finetuning_type": "full",
+    "reward_model": "",
+    "reward_model_type": "full",
+    "dataset": "system_chat",
+    "dataset_dir": "REMOTE:" + DEMO_DATA,
+    "template": "llama3",
+    "cutoff_len": 8192,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "fp16": True,
+}
+
+
+@pytest.mark.parametrize("num_samples", [16])
+def test_unsupervised_data(num_samples: int):
+    train_dataset = load_train_dataset(**TRAIN_ARGS)
+    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)
+    for index in indexes:
+        messages = original_data["messages"][index]
+        ref_ids = ref_tokenizer.apply_chat_template(messages)
+        ref_input_ids = ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True)
+        ref_labels = ref_ids[len(ref_input_ids) :]
+        assert train_dataset["input_ids"][index] == ref_input_ids
+        assert train_dataset["labels"][index] == ref_labels