LLaMA-Factory/tests/data/test_supervised.py

68 lines
2.4 KiB
Python
Raw Normal View History

2024-06-15 09:54:33 +00:00
# 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.
2024-06-07 21:20:54 +00:00
import os
2024-06-10 13:24:15 +00:00
import random
2024-06-07 21:20:54 +00:00
import pytest
from datasets import load_dataset
2024-06-19 12:03:50 +00:00
from transformers import AutoTokenizer
2024-06-07 21:20:54 +00:00
from llamafactory.data import get_dataset
from llamafactory.hparams import get_train_args
from llamafactory.model import load_tokenizer
2024-06-10 13:24:15 +00:00
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
2024-06-07 21:20:54 +00:00
2024-06-10 13:24:15 +00:00
TRAIN_ARGS = {
2024-06-07 21:20:54 +00:00
"model_name_or_path": TINY_LLAMA,
"stage": "sft",
"do_train": True,
"finetuning_type": "full",
2024-06-10 13:24:15 +00:00
"dataset": "llamafactory/tiny-supervised-dataset",
2024-06-07 21:20:54 +00:00
"dataset_dir": "ONLINE",
"template": "llama3",
2024-06-10 13:24:15 +00:00
"cutoff_len": 8192,
2024-06-07 21:20:54 +00:00
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
2024-06-15 11:51:20 +00:00
@pytest.mark.parametrize("num_samples", [16])
2024-06-10 13:24:15 +00:00
def test_supervised(num_samples: int):
model_args, data_args, training_args, _, _ = get_train_args(TRAIN_ARGS)
2024-06-07 21:20:54 +00:00
tokenizer_module = load_tokenizer(model_args)
tokenizer = tokenizer_module["tokenizer"]
tokenized_data = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
2024-06-19 12:03:50 +00:00
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
2024-06-10 13:24:15 +00:00
original_data = load_dataset(TRAIN_ARGS["dataset"], 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"]
2024-06-07 21:20:54 +00:00
messages = [
2024-06-10 13:24:15 +00:00
{"role": "user", "content": prompt},
{"role": "assistant", "content": original_data[index]["output"]},
2024-06-07 21:20:54 +00:00
]
2024-06-19 12:03:50 +00:00
templated_result = ref_tokenizer.apply_chat_template(messages, tokenize=False)
decoded_result = tokenizer.decode(tokenized_data["input_ids"][index])
assert templated_result == decoded_result