File size: 5,649 Bytes
83ddd7e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 | # 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.extras.packages import is_transformers_version_greater_than
from llamafactory.train.test_utils import load_dataset_module
DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
TRAIN_ARGS = {
"model_name_or_path": TINY_LLAMA3,
"stage": "sft",
"do_train": True,
"finetuning_type": "full",
"template": "llama3",
"cutoff_len": 8192,
"output_dir": "dummy_dir",
"overwrite_output_dir": True,
"fp16": True,
}
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [16])
def test_supervised_single_turn(num_samples: int):
train_dataset = load_dataset_module(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)["train_dataset"]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
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)
ref_prompt_ids = ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True)
if is_transformers_version_greater_than("5.0.0"):
ref_input_ids = ref_input_ids["input_ids"]
ref_prompt_ids = ref_prompt_ids["input_ids"]
prompt_len = len(ref_prompt_ids)
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
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [8])
def test_supervised_multi_turn(num_samples: int):
train_dataset = load_dataset_module(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)[
"train_dataset"
]
ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
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])
if is_transformers_version_greater_than("5.0.0"):
ref_input_ids = ref_input_ids["input_ids"]
# cannot test the label ids in multi-turn case
assert train_dataset["input_ids"][index] == ref_input_ids
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_train_on_prompt(num_samples: int):
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_LLAMA3)
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])
if is_transformers_version_greater_than("5.0.0"):
ref_input_ids = ref_input_ids["input_ids"]
assert train_dataset["input_ids"][index] == ref_input_ids
assert train_dataset["labels"][index] == ref_input_ids
@pytest.mark.runs_on(["cpu", "mps"])
@pytest.mark.parametrize("num_samples", [4])
def test_supervised_mask_history(num_samples: int):
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_LLAMA3)
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)
ref_prompt_ids = ref_tokenizer.apply_chat_template(messages[:-1], add_generation_prompt=True)
if is_transformers_version_greater_than("5.0.0"):
ref_input_ids = ref_input_ids["input_ids"]
ref_prompt_ids = ref_prompt_ids["input_ids"]
prompt_len = len(ref_prompt_ids)
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
|