tests/data/test_base.py

# Copyright Lightning AI. Licensed under the Apache License 2.0, see LICENSE file.

from typing import Optional

import pytest
import torch

from litgpt.data.base import SFTDataset, get_sft_collate_fn
from litgpt.prompts import PromptStyle


@pytest.mark.parametrize("mask_prompt", [True, False])
@pytest.mark.parametrize("ignore_index", [-1, -100])
@pytest.mark.parametrize("max_seq_length", [1000, 5, -1])
def test_sft_dataset(max_seq_length, ignore_index, mask_prompt, mock_tokenizer):
    class Style(PromptStyle):
        def apply(self, prompt: str, *, sys_prompt: Optional[str] = None, **kwargs) -> str:
            return f"In: {prompt} Out:"

    i = ignore_index
    data = [{"instruction": "Foo", "output": "Bar"}, {"instruction": "Boo", "output": "Ahh"}]

    dataset = SFTDataset(
        data=data,
        tokenizer=mock_tokenizer,
        prompt_style=Style(),
        mask_prompt=mask_prompt,
        ignore_index=ignore_index,
        max_seq_length=max_seq_length,
    )
    assert len(dataset) == len(data)

    expected_input_ids = torch.tensor([73, 110, 58, 32, 70, 111, 111, 32, 79, 117, 116, 58, 66, 97, 114, 1])
    # If prompt is not masked, labels == input_ids
    expected_labels = (
        torch.tensor([i, i, i, i, i, i, i, i, i, i, i, i, 66, 97, 114, 1]) if mask_prompt else expected_input_ids
    )

    if max_seq_length == -1:
        assert torch.equal(dataset[0]["input_ids"], expected_input_ids)
        assert torch.equal(dataset[0]["labels"], expected_labels)
    else:
        assert torch.equal(dataset[0]["input_ids"], expected_input_ids[:max_seq_length])
        assert torch.equal(dataset[0]["labels"], expected_labels[:max_seq_length])


@pytest.mark.parametrize("ignore_index", [-1, -100])
@pytest.mark.parametrize("pad_id", [0, 100])
def test_sft_collate_fn_padding(pad_id, ignore_index):
    collate = get_sft_collate_fn(pad_id=pad_id, ignore_index=ignore_index)
    samples = [
        {
            "input_ids": torch.tensor([1, 2, 3]),
            "labels": torch.tensor([10, 20, 30]),
            "token_counts": {"raw": 3, "raw_plus_prompt_template": 25},
        },
        {
            "input_ids": torch.tensor([4, 5, 6, 7, 8]),
            "labels": torch.tensor([40, 50, 60, 70, 80]),
            "token_counts": {"raw": 5, "raw_plus_prompt_template": 27},
        },
    ]
    expected = {
        "input_ids": torch.tensor([[1, 2, 3, pad_id, pad_id], [4, 5, 6, 7, 8]]),
        "labels": torch.tensor([[10, 20, 30, ignore_index, ignore_index], [40, 50, 60, 70, 80]]),
        "token_counts": {"raw": torch.tensor([[3], [5]]), "raw_plus_prompt_template": torch.tensor([[25], [27]])},
    }
    batch = collate(samples)
    assert all(torch.equal(batch[k], expected[k]) for k in ("input_ids", "labels"))
    for key in ("raw", "raw_plus_prompt_template"):
        assert torch.equal(batch["token_counts"][key], expected["token_counts"][key]), f"Token count mismatch for {key}"


def test_sft_collate_fn_truncation():
    collate = get_sft_collate_fn(max_seq_length=2)
    samples = [
        {
            "input_ids": torch.tensor([1, 2, 3]),
            "labels": torch.tensor([10, 20, 30]),
            "token_counts": {"raw": 3, "raw_plus_prompt_template": 25},
        },
        {
            "input_ids": torch.tensor([4, 5, 6, 7, 8]),
            "labels": torch.tensor([40, 50, 60, 70, 80]),
            "token_counts": {"raw": 5, "raw_plus_prompt_template": 27},
        },
    ]
    expected = {
        "input_ids": torch.tensor([[1, 2], [4, 5]]),
        "labels": torch.tensor([[10, 20], [40, 50]]),
        "token_counts": {"raw": torch.tensor([[3], [5]]), "raw_plus_prompt_template": torch.tensor([[25], [27]])},
    }
    batch = collate(samples)
    assert all(torch.equal(batch[k], expected[k]) for k in ("input_ids", "labels"))
    for key in ("raw", "raw_plus_prompt_template"):
        assert torch.equal(batch["token_counts"][key], expected["token_counts"][key]), f"Token count mismatch for {key}"