File size: 1,733 Bytes
86e3bb9 | 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 | from __future__ import annotations
import numpy as np
import torch
from kawk50m.chat import build_assistant_examples, format_generation_prompt, parse_chatml
from kawk50m.sft_dataset import SFTPackedDataset
class CharacterTokenizer:
bos_token_id = 1
eos_token_id = 2
def __call__(self, text: str, **_: object) -> dict[str, list[int]]:
return {"input_ids": [10 + ord(character) % 200 for character in text]}
def test_chatml_examples_mask_only_current_assistant_response() -> None:
blocks = parse_chatml(
"<|im_start|>user\n질문입니다.<|im_end|>"
"<|im_start|>assistant\n답변입니다.<|im_end|>"
)
examples = list(
build_assistant_examples(blocks, CharacterTokenizer(), max_length=128)
)
assert len(examples) == 1
input_ids, loss_mask = examples[0]
assert len(input_ids) == len(loss_mask)
assert loss_mask[-1] == 1
assert input_ids[-1] == CharacterTokenizer.eos_token_id
assert 0 in loss_mask and 1 in loss_mask
assert format_generation_prompt("질문") .endswith("### 도우미:\n")
def test_sft_dataset_applies_ignore_index(tmp_path) -> None:
prefix = tmp_path / "train"
np.asarray([1, 10, 11, 2, 1, 20, 21, 2], dtype=np.uint16).tofile(
str(prefix) + ".input_ids.bin"
)
np.asarray([0, 0, 1, 1, 0, 0, 1, 1], dtype=np.uint8).tofile(
str(prefix) + ".loss_mask.bin"
)
dataset = SFTPackedDataset(prefix, sequence_length=4)
inputs, labels = dataset.evaluation_batch(
batch_index=0,
batch_size=1,
rank=0,
world_size=1,
device=torch.device("cpu"),
)
assert inputs.tolist() == [[1, 10, 11, 2]]
assert labels.tolist() == [[-100, -100, 11, 2]]
|