"""
Preprocess OpenThoughts3-1.2M (code domain only).

Source: https://huggingface.co/datasets/open-thoughts/OpenThoughts3-1.2M

This script filters the code domain, deduplicates by user prompt
(picking one random rollout per prompt via reservoir sampling),
asserts correctness, and uploads to HuggingFace.
"""

from __future__ import annotations

import json
import random
import tempfile
from pathlib import Path

from datasets import load_dataset
from huggingface_hub import HfApi, create_repo


DATASET_ID = "open-thoughts/OpenThoughts3-1.2M"
EXPECTED_UNIQUE_PROMPTS = 5_714
HF_REPO_ID = "teetone/openthoughts3-code-deduped"


def extract_prompt(conversations: list[dict]) -> str | None:
    """Return the first human turn from a conversation."""
    for turn in conversations:
        if turn["from"] == "human":
            return turn["value"]
    return None


def preprocess_and_upload(seed: int = 42) -> None:
    random.seed(seed)

    print(f"Loading {DATASET_ID} (streaming) ...")
    ds = load_dataset(DATASET_ID, split="train", streaming=True)

    # Single streaming pass: reservoir-sample one row per unique prompt.
    # For each prompt we track (chosen_row, count_seen) so we can do
    # uniform random selection without storing all rollouts.
    prompt_reservoir: dict[str, tuple[dict, int]] = {}
    total = 0
    code_count = 0

    for row in ds:
        total += 1
        if row["domain"] != "code":
            continue
        code_count += 1

        prompt = extract_prompt(row["conversations"])
        if prompt is None:
            continue

        rec = {
            "difficulty": row["difficulty"],
            "source": row["source"],
            "domain": row["domain"],
            "conversations": row["conversations"],
        }

        if prompt not in prompt_reservoir:
            prompt_reservoir[prompt] = (rec, 1)
        else:
            _, count = prompt_reservoir[prompt]
            count += 1
            # Reservoir sampling: replace with probability 1/count
            if random.randint(1, count) == 1:
                prompt_reservoir[prompt] = (rec, count)
            else:
                prompt_reservoir[prompt] = (prompt_reservoir[prompt][0], count)

        if code_count % 100_000 == 0:
            print(f"  Processed {code_count:,} code rows, {len(prompt_reservoir):,} unique prompts ...")

    print(f"  Total rows streamed: {total:,}")
    print(f"  Code rows: {code_count:,}")
    print(f"  Unique prompts: {len(prompt_reservoir):,}")

    records = [row for row, _ in prompt_reservoir.values()]
    print(f"  Deduplicated rows: {len(records):,}")

    # ── Assertions ────────────────────────────────────────────────────
    assert len(records) == EXPECTED_UNIQUE_PROMPTS, (
        f"Expected {EXPECTED_UNIQUE_PROMPTS} rows, got {len(records)}"
    )

    seen_prompts: set[str] = set()
    for rec in records:
        assert rec["domain"] == "code", f"Non-code row found: domain={rec['domain']}"
        prompt = extract_prompt(rec["conversations"])
        assert prompt not in seen_prompts, f"Duplicate prompt found: {prompt[:80]}..."
        seen_prompts.add(prompt)

    assert len(seen_prompts) == EXPECTED_UNIQUE_PROMPTS, (
        f"Expected {EXPECTED_UNIQUE_PROMPTS} unique prompts, got {len(seen_prompts)}"
    )
    print("  All assertions passed.")

    # ── Upload to HuggingFace ─────────────────────────────────────────
    print(f"Uploading to {HF_REPO_ID} ...")
    api = HfApi()
    create_repo(HF_REPO_ID, repo_type="dataset", exist_ok=True)

    with tempfile.NamedTemporaryFile(mode="w", suffix=".jsonl", delete=False) as f:
        temp_path = f.name
        for rec in records:
            f.write(json.dumps(rec, ensure_ascii=False) + "\n")

    api.upload_file(
        path_or_fileobj=temp_path,
        path_in_repo="data/train-00000-of-00001.jsonl",
        repo_id=HF_REPO_ID,
        repo_type="dataset",
    )
    Path(temp_path).unlink()

    print(f"Done! Dataset available at: https://huggingface.co/datasets/{HF_REPO_ID}")


if __name__ == "__main__":
    preprocess_and_upload()