"""Run from the MiniMind repository root with its training dependencies installed."""

import json
import random
import sys
from pathlib import Path
from tempfile import TemporaryDirectory

from transformers import AutoTokenizer

sys.path.insert(0, str(Path.cwd()))
from dataset.lm_dataset import SFTDataset


def sample(repeats):
    return {"conversations": [
        {"role": "system", "content": "Test."},
        {"role": "user", "content": "status " * repeats},
        {"role": "assistant", "content": "OK"},
    ]}


def main():
    tokenizer = AutoTokenizer.from_pretrained("./model", local_files_only=True)
    repeats = (1, 500)
    with TemporaryDirectory() as directory:
        path = Path(directory) / "samples.jsonl"
        path.write_text("\n".join(json.dumps(sample(n)) for n in repeats) + "\n", encoding="utf-8")
        dataset = SFTDataset(str(path), tokenizer, max_length=768)
        print("prompt_repeats input_slots supervised_labels")
        for index, count in enumerate(repeats):
            random.seed(42)  # Make the loader's empty thinking-tag treatment reproducible.
            inputs, labels = dataset[index]
            print(count, len(inputs), (labels != -100).sum().item())


if __name__ == "__main__":
    main()
