ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix #2666

Browse Source 
Former-commit-id: f121d5c4f94af9f165132c4309cb9bdc8217d985
This commit is contained in:
hiyouga
2024-06-10 21:24:15 +08:00
parent 0ecf0d51e3
commit 784088db3f
6 changed files with 41 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
tests/data/test_supervised.py
View File

@@ -1,4 +1,5 @@
import os
import random
import pytest
from datasets import load_dataset
@@ -8,17 +9,17 @@ from llamafactory.hparams import get_train_args
from llamafactory.model import load_tokenizer
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-LlamaForCausalLM")
TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
TRAINING_ARGS = {
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA,
"stage": "sft",
"do_train": True,
"finetuning_type": "full",
"dataset": "llamafactory/tiny_dataset",
"dataset": "llamafactory/tiny-supervised-dataset",
"dataset_dir": "ONLINE",
"template": "llama3",
"cutoff_len": 1024,
"cutoff_len": 8192,
"overwrite_cache": True,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
@@ -26,19 +27,24 @@ TRAINING_ARGS = {
}
@pytest.mark.parametrize("test_num", [5])
def test_supervised(test_num: int):
model_args, data_args, training_args, _, _ = get_train_args(TRAINING_ARGS)
@pytest.mark.parametrize("num_samples", [10])
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)
original_data = load_dataset(TRAINING_ARGS["dataset"], split="train")
for test_idx in range(test_num):
decode_result = tokenizer.decode(tokenized_data["input_ids"][test_idx])
original_data = load_dataset(TRAIN_ARGS["dataset"], split="train")
indexes = random.choices(range(len(original_data)), k=num_samples)
for index in indexes:
decoded_result = tokenizer.decode(tokenized_data["input_ids"][index])
prompt = original_data[index]["instruction"]
if original_data[index]["input"]:
prompt += "\n" + original_data[index]["input"]
messages = [
{"role": "user", "content": original_data[test_idx]["instruction"]},
{"role": "assistant", "content": original_data[test_idx]["output"]},
{"role": "user", "content": prompt},
{"role": "assistant", "content": original_data[index]["output"]},
]
templated_result = tokenizer.apply_chat_template(messages, tokenize=False)
assert decode_result == templated_result
assert decoded_result == templated_result