add unittest

Former-commit-id: 8a1f0c5f922989e08a19c65de0b2c4afd2a5771f

tests/data/processors/test_feedback.py

@@ -19,20 +19,21 @@ import pytest
 from datasets import load_dataset
 from transformers import AutoTokenizer
 
-from llamafactory.data import get_dataset
-from llamafactory.hparams import get_train_args
-from llamafactory.model import load_tokenizer
+from llamafactory.extras.constants import IGNORE_INDEX
+from llamafactory.train.test_utils import load_train_dataset
 
 
+DEMO_DATA = os.environ.get("DEMO_DATA", "llamafactory/demo_data")
+
 TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
 
 TRAIN_ARGS = {
     "model_name_or_path": TINY_LLAMA,
-    "stage": "sft",
+    "stage": "kto",
     "do_train": True,
     "finetuning_type": "full",
-    "dataset": "llamafactory/tiny-supervised-dataset",
-    "dataset_dir": "ONLINE",
+    "dataset": "kto_en_demo",
+    "dataset_dir": "REMOTE:" + DEMO_DATA,
     "template": "llama3",
     "cutoff_len": 8192,
     "overwrite_cache": True,
@@ -43,25 +44,16 @@ TRAIN_ARGS = {
 
 
 @pytest.mark.parametrize("num_samples", [16])
-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"]
-    dataset_module = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
-
+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(TRAIN_ARGS["dataset"], split="train")
+    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:
-        prompt = original_data[index]["instruction"]
-        if original_data[index]["input"]:
-            prompt += "\n" + original_data[index]["input"]
-
-        messages = [
-            {"role": "user", "content": prompt},
-            {"role": "assistant", "content": original_data[index]["output"]},
-        ]
-        templated_result = ref_tokenizer.apply_chat_template(messages, tokenize=False)
-        decoded_result = tokenizer.decode(dataset_module["train_dataset"]["input_ids"][index])
-        assert templated_result == decoded_result
+        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/processors/test_pairwise.py

@@ -0,0 +1,78 @@
+# Copyright 2024 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.environ.get("DEMO_DATA", "llamafactory/demo_data")
+
+TINY_LLAMA = os.environ.get("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/processors/test_supervised.py

@@ -0,0 +1,104 @@
+# Copyright 2024 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.environ.get("DEMO_DATA", "llamafactory/demo_data")
+
+TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+TINY_DATA = os.environ.get("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/processors/test_unsupervised.py

@@ -0,0 +1,58 @@
+# Copyright 2024 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.environ.get("DEMO_DATA", "llamafactory/demo_data")
+
+TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+TINY_DATA = os.environ.get("TINY_DATA", "llamafactory/tiny-supervised-dataset")
+
+TRAIN_ARGS = {
+    "model_name_or_path": TINY_LLAMA,
+    "stage": "sft",
+    "do_predict": True,
+    "finetuning_type": "full",
+    "template": "llama3",
+    "cutoff_len": 8192,
+    "overwrite_cache": True,
+    "output_dir": "dummy_dir",
+    "overwrite_output_dir": True,
+    "predict_with_generate": True,
+    "fp16": True,
+}
+
+
+@pytest.mark.parametrize("num_samples", [16])
+def test_unsupervised_data(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:
+        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

tests/data/test_template.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import os
-from typing import TYPE_CHECKING, Sequence
+from typing import TYPE_CHECKING, List, Sequence
 
 import pytest
 from transformers import AutoTokenizer
@@ -39,7 +39,7 @@ MESSAGES = [
 
 def _check_tokenization(
     tokenizer: "PreTrainedTokenizer", batch_input_ids: Sequence[Sequence[int]], batch_text: Sequence[str]
-):
+) -> None:
     for input_ids, text in zip(batch_input_ids, batch_text):
         assert input_ids == tokenizer.encode(text, add_special_tokens=False)
         assert tokenizer.decode(input_ids) == text
@@ -47,7 +47,7 @@ def _check_tokenization(
 
 def _check_single_template(
     model_id: str, template_name: str, prompt_str: str, answer_str: str, extra_str: str, use_fast: bool
-):
+) -> List[str]:
     tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=use_fast, token=HF_TOKEN)
     content_str = tokenizer.apply_chat_template(MESSAGES, tokenize=False)
     content_ids = tokenizer.apply_chat_template(MESSAGES, tokenize=True)
@@ -59,7 +59,7 @@ def _check_single_template(
     return content_ids
 
 
-def _check_template(model_id: str, template_name: str, prompt_str: str, answer_str: str, extra_str: str = ""):
+def _check_template(model_id: str, template_name: str, prompt_str: str, answer_str: str, extra_str: str = "") -> None:
     """
     Checks template for both the slow tokenizer and the fast tokenizer.