scripts/build_general_science_release.py

#!/usr/bin/env python3
"""Build the general science QA release before upload.

The script reads local parquet snapshots only. It creates train/test JSONL
files, exports ScienceQA images, computes release statistics, and writes a
dataset card.
"""

from __future__ import annotations

import hashlib
import json
import random
import re
import sys
from collections import Counter, defaultdict
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Callable

try:
    import pyarrow.parquet as pq
except ImportError as exc:
    raise SystemExit("pyarrow is required. Run with: conda run -n memory python ...") from exc


RELEASE_DIR = Path(__file__).resolve().parents[1]
DATA_ROOT = RELEASE_DIR.parent
IMAGE_DIR = RELEASE_DIR / "images" / "scienceqa"
TRAIN_PATH = RELEASE_DIR / "train.jsonl"
TEST_PATH = RELEASE_DIR / "test.jsonl"
STATS_PATH = RELEASE_DIR / "stats.json"
CARD_PATH = RELEASE_DIR / "general_science_card.md"
README_PATH = RELEASE_DIR / "README.md"

SEED = 42
TASK_TYPE = "multiple_choice_science_qa"

SOURCE_LICENSES = {
    "allenai/sciq": {
        "license": "CC BY-NC 3.0",
        "license_id": "cc-by-nc-3.0",
        "license_source": "local Hugging Face dataset card: sciq/README.md",
        "url": "https://huggingface.co/datasets/allenai/sciq",
    },
    "allenai/ai2_arc": {
        "license": "CC BY-SA 4.0",
        "license_id": "cc-by-sa-4.0",
        "license_source": "local Hugging Face dataset card: ai2_arc/README.md",
        "url": "https://huggingface.co/datasets/allenai/ai2_arc",
    },
    "derek-thomas/ScienceQA": {
        "license": "TBD: not available in local snapshot",
        "license_id": "unknown-local-snapshot",
        "license_source": "local parquet snapshot does not include README/license metadata; verify upstream before public upload",
        "url": "https://huggingface.co/datasets/derek-thomas/ScienceQA",
    },
}


@dataclass(frozen=True)
class SourceFile:
    display_name: str
    dataset: str
    subset: str | None
    source_split: str
    final_split: str
    path: Path
    converter: Callable[[dict[str, Any], int, "SourceFile"], dict[str, Any] | None]


def read_rows(path: Path) -> list[dict[str, Any]]:
    if not path.exists():
        raise FileNotFoundError(f"Required parquet file not found: {path}")
    return pq.read_table(path).to_pylist()


def find_single_parquet(pattern: str) -> Path:
    matches = sorted(DATA_ROOT.glob(pattern))
    if len(matches) != 1:
        found = ", ".join(str(path.relative_to(DATA_ROOT)) for path in matches) or "none"
        raise FileNotFoundError(f"Expected exactly one parquet for {pattern}; found {found}")
    return matches[0]


def clean_text(value: Any) -> str | None:
    if value is None:
        return None
    text = str(value).strip()
    return text or None


def normalize_for_key(value: Any) -> str:
    text = clean_text(value) or ""
    text = text.lower()
    text = re.sub(r"\s+", " ", text)
    text = re.sub(r"[^\w\s]+", "", text)
    return text.strip()


def generated_label(index: int) -> str:
    label = ""
    index += 1
    while index:
        index, remainder = divmod(index - 1, 26)
        label = chr(ord("A") + remainder) + label
    return label


def stable_rng(key: str) -> random.Random:
    digest = hashlib.sha256(f"{SEED}:{key}".encode("utf-8")).hexdigest()
    return random.Random(int(digest[:16], 16))


def sanitize_id(value: str) -> str:
    return re.sub(r"[^A-Za-z0-9_.:-]+", "-", value).strip("-")


def detect_image_format(image_bytes: bytes) -> tuple[str, str]:
    if image_bytes.startswith(b"\x89PNG\r\n\x1a\n"):
        return "png", "image/png"
    if image_bytes.startswith(b"\xff\xd8\xff"):
        return "jpg", "image/jpeg"
    if image_bytes.startswith(b"GIF87a") or image_bytes.startswith(b"GIF89a"):
        return "gif", "image/gif"
    if image_bytes.startswith(b"RIFF") and image_bytes[8:12] == b"WEBP":
        return "webp", "image/webp"
    return "bin", "application/octet-stream"


def extract_image_payload(row: dict[str, Any]) -> dict[str, Any] | None:
    image = row.get("image")
    if not isinstance(image, dict):
        return None

    image_bytes = image.get("bytes")
    original_path = clean_text(image.get("path"))
    if isinstance(image_bytes, bytes) and image_bytes:
        extension, mime_type = detect_image_format(image_bytes)
        return {
            "bytes": image_bytes,
            "extension": extension,
            "mime_type": mime_type,
            "original_path": original_path,
        }
    if original_path:
        return {
            "bytes": None,
            "extension": None,
            "mime_type": None,
            "original_path": original_path,
        }
    return None


def materialize_image(example_id: str, payload: dict[str, Any] | None) -> dict[str, Any] | None:
    if payload is None:
        return None

    image_bytes = payload.get("bytes")
    if isinstance(image_bytes, bytes) and image_bytes:
        IMAGE_DIR.mkdir(parents=True, exist_ok=True)
        extension = payload["extension"]
        image_path = IMAGE_DIR / f"{example_id}.{extension}"
        image_path.write_bytes(image_bytes)
        return {
            "path": str(image_path.relative_to(RELEASE_DIR)),
            "mime_type": payload["mime_type"],
        }

    original_path = payload.get("original_path")
    if original_path:
        return {"path": original_path, "mime_type": payload.get("mime_type")}
    return None


def combine_support(*parts: Any) -> str | None:
    cleaned = [text for text in (clean_text(part) for part in parts) if text]
    return "\n\n".join(cleaned) if cleaned else None


def make_dedup_key(question: str, choices: list[dict[str, Any]], answer_text: str) -> str:
    choice_texts = sorted(normalize_for_key(choice["text"]) for choice in choices)
    return " || ".join([normalize_for_key(question), *choice_texts, normalize_for_key(answer_text)])


def shuffle_and_relabel(raw_choices: list[dict[str, Any]], example_id: str) -> tuple[list[dict[str, str]], str | None]:
    choices = list(raw_choices)
    stable_rng(example_id).shuffle(choices)

    output_choices: list[dict[str, str]] = []
    answer_label = None
    for index, choice in enumerate(choices):
        label = generated_label(index)
        output_choices.append({"label": label, "text": choice["text"]})
        if choice.get("is_correct"):
            answer_label = label
    return output_choices, answer_label


def build_example(item: dict[str, Any]) -> dict[str, Any]:
    choices, answer_label = shuffle_and_relabel(item["raw_choices"], item["id"])
    image = materialize_image(item["id"], item.get("image_payload"))
    modality = "image_text" if image else "text"

    example = {
        "id": item["id"],
        "dataset": item["dataset"],
        "subset": item["subset"],
        "split": item["split"],
        "task_type": TASK_TYPE,
        "modality": modality,
        "question": item["question"],
        "image": image,
        "choices": choices,
        "answer_label": answer_label,
        "answer_text": item["answer_text"],
        "support": item["support"],
        "source_meta": item["source_meta"],
    }
    validate_example(example)
    return example


def validate_example(example: dict[str, Any]) -> None:
    required = [
        "id",
        "dataset",
        "split",
        "task_type",
        "modality",
        "question",
        "image",
        "choices",
        "answer_label",
        "answer_text",
        "source_meta",
    ]
    missing = [field for field in required if field not in example]
    if missing:
        raise ValueError(f"{example.get('id')} missing fields: {missing}")
    if not example["question"] or not example["answer_label"] or not example["answer_text"]:
        raise ValueError(f"{example['id']} has empty question or answer")
    if not isinstance(example["choices"], list) or len(example["choices"]) < 4:
        raise ValueError(f"{example['id']} has fewer than 4 choices")
    if example["answer_label"] not in {choice["label"] for choice in example["choices"]}:
        raise ValueError(f"{example['id']} answer_label is absent from choices")
    if example["modality"] == "image_text" and not example["image"]:
        raise ValueError(f"{example['id']} image_text example has no image")


def convert_sciq(row: dict[str, Any], index: int, source: SourceFile) -> dict[str, Any] | None:
    question = clean_text(row.get("question"))
    answer_text = clean_text(row.get("correct_answer"))
    support = clean_text(row.get("support"))
    choice_texts = [
        clean_text(row.get("distractor1")),
        clean_text(row.get("distractor2")),
        clean_text(row.get("distractor3")),
        answer_text,
    ]
    if question is None or answer_text is None or any(choice is None for choice in choice_texts):
        return None

    raw_choices = [
        {"text": choice, "is_correct": idx == 3}
        for idx, choice in enumerate(choice_texts)
        if choice is not None
    ]
    if len(raw_choices) < 4:
        return None

    example_id = f"sciq-{source.source_split}-{index:05d}"
    return {
        "id": example_id,
        "dataset": source.dataset,
        "subset": source.subset,
        "split": source.final_split,
        "question": question,
        "raw_choices": raw_choices,
        "answer_text": answer_text,
        "support": support,
        "image_payload": None,
        "source_meta": {
            "source_file": str(source.path.relative_to(DATA_ROOT)),
            "source_split": source.source_split,
            "source_index": index,
            "original_answer_text": answer_text,
            **SOURCE_LICENSES[source.dataset],
        },
    }


def extract_struct_choices(raw_choices: Any) -> list[dict[str, str]] | None:
    if not isinstance(raw_choices, dict):
        return None
    texts = raw_choices.get("text")
    labels = raw_choices.get("label")
    if not isinstance(texts, list) or not isinstance(labels, list) or len(texts) != len(labels):
        return None
    choices: list[dict[str, str]] = []
    for label, text in zip(labels, texts):
        clean_label = clean_text(label)
        clean_choice_text = clean_text(text)
        if clean_label is None or clean_choice_text is None:
            return None
        choices.append({"label": clean_label, "text": clean_choice_text})
    return choices


def convert_ai2_arc(row: dict[str, Any], index: int, source: SourceFile) -> dict[str, Any] | None:
    question = clean_text(row.get("question"))
    answer_key = clean_text(row.get("answerKey"))
    choices = extract_struct_choices(row.get("choices"))
    if question is None or answer_key is None or choices is None or len(choices) < 4:
        return None

    answer_text = None
    for choice in choices:
        if choice["label"] == answer_key or choice["label"].lower() == answer_key.lower():
            answer_text = choice["text"]
            break
    if answer_text is None:
        return None

    raw_choices = [
        {"text": choice["text"], "is_correct": choice["text"] == answer_text}
        for choice in choices
    ]
    original_id = clean_text(row.get("id")) or f"{source.source_split}-{index:05d}"
    subset_id = sanitize_id(source.subset or "default")
    example_id = f"ai2_arc-{subset_id}-{source.source_split}-{sanitize_id(original_id)}"
    return {
        "id": example_id,
        "dataset": source.dataset,
        "subset": source.subset,
        "split": source.final_split,
        "question": question,
        "raw_choices": raw_choices,
        "answer_text": answer_text,
        "support": None,
        "image_payload": None,
        "source_meta": {
            "source_file": str(source.path.relative_to(DATA_ROOT)),
            "source_split": source.source_split,
            "source_index": index,
            "original_id": original_id,
            "original_answer_label": answer_key,
            "original_choice_labels": [choice["label"] for choice in choices],
            **SOURCE_LICENSES[source.dataset],
        },
    }


def convert_scienceqa(row: dict[str, Any], index: int, source: SourceFile) -> dict[str, Any] | None:
    question = clean_text(row.get("question"))
    raw_choice_values = row.get("choices")
    answer_index = row.get("answer")
    subject = clean_text(row.get("subject"))
    if subject != "natural science":
        return None
    if question is None or not isinstance(raw_choice_values, list):
        return None
    choice_texts = [clean_text(choice) for choice in raw_choice_values]
    if any(choice is None for choice in choice_texts) or len(choice_texts) < 4:
        return None
    if not isinstance(answer_index, int) or isinstance(answer_index, bool):
        return None
    if answer_index < 0 or answer_index >= len(choice_texts):
        return None

    answer_text = choice_texts[answer_index]
    raw_choices = [
        {"text": choice, "is_correct": choice_index == answer_index}
        for choice_index, choice in enumerate(choice_texts)
        if choice is not None
    ]
    example_id = f"scienceqa-{source.source_split}-{index:05d}"
    image_payload = extract_image_payload(row)
    return {
        "id": example_id,
        "dataset": source.dataset,
        "subset": source.subset,
        "split": source.final_split,
        "question": question,
        "raw_choices": raw_choices,
        "answer_text": answer_text,
        "support": combine_support(row.get("hint"), row.get("lecture"), row.get("solution")),
        "image_payload": image_payload,
        "source_meta": {
            "source_file": str(source.path.relative_to(DATA_ROOT)),
            "source_split": source.source_split,
            "source_index": index,
            "original_answer_index": answer_index,
            "task": clean_text(row.get("task")),
            "grade": clean_text(row.get("grade")),
            "subject": subject,
            "topic": clean_text(row.get("topic")),
            "category": clean_text(row.get("category")),
            "skill": clean_text(row.get("skill")),
            "image_present": image_payload is not None,
            "original_image_path": image_payload.get("original_path") if image_payload else None,
            **SOURCE_LICENSES[source.dataset],
        },
    }


def load_source_files() -> list[SourceFile]:
    scienceqa_train = find_single_parquet("scienceqa_hf/data/train-*.parquet")
    scienceqa_validation = find_single_parquet("scienceqa_hf/data/validation-*.parquet")
    scienceqa_test = find_single_parquet("scienceqa_hf/data/test-*.parquet")

    files: list[SourceFile] = [
        SourceFile("SciQ", "allenai/sciq", None, "train", "train", DATA_ROOT / "sciq/data/train-00000-of-00001.parquet", convert_sciq),
        SourceFile("SciQ", "allenai/sciq", None, "validation", "train", DATA_ROOT / "sciq/data/validation-00000-of-00001.parquet", convert_sciq),
        SourceFile("SciQ", "allenai/sciq", None, "test", "test", DATA_ROOT / "sciq/data/test-00000-of-00001.parquet", convert_sciq),
        SourceFile("AI2 ARC-Challenge", "allenai/ai2_arc", "ARC-Challenge", "train", "train", DATA_ROOT / "ai2_arc/ARC-Challenge/train-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("AI2 ARC-Challenge", "allenai/ai2_arc", "ARC-Challenge", "validation", "train", DATA_ROOT / "ai2_arc/ARC-Challenge/validation-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("AI2 ARC-Challenge", "allenai/ai2_arc", "ARC-Challenge", "test", "test", DATA_ROOT / "ai2_arc/ARC-Challenge/test-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("AI2 ARC-Easy", "allenai/ai2_arc", "ARC-Easy", "train", "train", DATA_ROOT / "ai2_arc/ARC-Easy/train-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("AI2 ARC-Easy", "allenai/ai2_arc", "ARC-Easy", "validation", "train", DATA_ROOT / "ai2_arc/ARC-Easy/validation-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("AI2 ARC-Easy", "allenai/ai2_arc", "ARC-Easy", "test", "test", DATA_ROOT / "ai2_arc/ARC-Easy/test-00000-of-00001.parquet", convert_ai2_arc),
        SourceFile("ScienceQA", "derek-thomas/ScienceQA", None, "train", "train", scienceqa_train, convert_scienceqa),
        SourceFile("ScienceQA", "derek-thomas/ScienceQA", None, "validation", "train", scienceqa_validation, convert_scienceqa),
        SourceFile("ScienceQA", "derek-thomas/ScienceQA", None, "test", "test", scienceqa_test, convert_scienceqa),
    ]
    return files


def collect_items() -> tuple[list[dict[str, Any]], dict[str, Any]]:
    raw_items: list[dict[str, Any]] = []
    source_stats: dict[str, Any] = {}

    for source in load_source_files():
        rows = read_rows(source.path)
        converted = 0
        skipped = 0
        for index, row in enumerate(rows):
            item = source.converter(row, index, source)
            if item is None:
                skipped += 1
                continue
            item["dedup_key"] = make_dedup_key(item["question"], item["raw_choices"], item["answer_text"])
            raw_items.append(item)
            converted += 1
        key = f"{source.dataset}|{source.subset}|{source.source_split}"
        source_stats[key] = {
            "display_name": source.display_name,
            "dataset": source.dataset,
            "subset": source.subset,
            "source_split": source.source_split,
            "final_split": source.final_split,
            "source_file": str(source.path.relative_to(DATA_ROOT)),
            "raw_rows": len(rows),
            "converted": converted,
            "skipped": skipped,
        }
        print(f"{source.display_name} {source.source_split}: converted {converted}, skipped {skipped}")
    return raw_items, source_stats


def deduplicate(items: list[dict[str, Any]]) -> tuple[list[dict[str, Any]], dict[str, Any]]:
    train = [item for item in items if item["split"] == "train"]
    test = [item for item in items if item["split"] == "test"]

    def dedup_split(split_items: list[dict[str, Any]]) -> tuple[list[dict[str, Any]], int]:
        seen: set[str] = set()
        output: list[dict[str, Any]] = []
        duplicates = 0
        for item in split_items:
            key = item["dedup_key"]
            if key in seen:
                duplicates += 1
                continue
            seen.add(key)
            output.append(item)
        return output, duplicates

    test, test_dups = dedup_split(test)
    test_keys = {item["dedup_key"] for item in test}
    train, train_dups = dedup_split(train)
    train_before_leak_filter = len(train)
    train = [item for item in train if item["dedup_key"] not in test_keys]
    train_removed_for_test_overlap = train_before_leak_filter - len(train)

    stats = {
        "train_duplicates_removed": train_dups,
        "test_duplicates_removed": test_dups,
        "train_removed_for_test_overlap": train_removed_for_test_overlap,
        "train_test_overlap_after_filter": len({item["dedup_key"] for item in train} & test_keys),
    }
    return train + test, stats


TOKEN_PATTERN = re.compile(r"\w+|[^\w\s]", re.UNICODE)


def get_token_counter() -> tuple[str, Callable[[str], int]]:
    try:
        import tiktoken  # type: ignore

        encoding = tiktoken.get_encoding("cl100k_base")
        return "cl100k_base via tiktoken", lambda text: len(encoding.encode(text or ""))
    except Exception:
        return "regex_approx_v1", lambda text: len(TOKEN_PATTERN.findall(text or ""))


def input_text(example: dict[str, Any]) -> str:
    choices = "\n".join(f"{choice['label']}. {choice['text']}" for choice in example["choices"])
    image_hint = f"\n[image: {example['image']['path']}]" if example.get("image") else ""
    return f"{example['question']}{image_hint}\n{choices}"


def compute_stats(examples: list[dict[str, Any]], source_stats: dict[str, Any], dedup_stats: dict[str, Any]) -> dict[str, Any]:
    tokenizer_name, count_tokens = get_token_counter()

    def empty_bucket() -> dict[str, Any]:
        return {
            "num_examples": 0,
            "input_tokens": 0,
            "support_tokens": 0,
            "full_record_tokens": 0,
            "image_examples": 0,
            "text_examples": 0,
        }

    buckets: dict[str, dict[str, Any]] = defaultdict(empty_bucket)
    by_dataset: dict[str, dict[str, Any]] = defaultdict(empty_bucket)
    by_modality = Counter()

    for example in examples:
        serialized = json.dumps(example, ensure_ascii=False, sort_keys=True)
        support = example.get("support") or ""
        values = {
            "num_examples": 1,
            "input_tokens": count_tokens(input_text(example)),
            "support_tokens": count_tokens(support),
            "full_record_tokens": count_tokens(serialized),
            "image_examples": 1 if example.get("image") else 0,
            "text_examples": 0 if example.get("image") else 1,
        }
        keys = [
            "overall",
            f"split::{example['split']}",
        ]
        dataset_key = f"{example['dataset']}|{example.get('subset')}"
        by_modality[example["modality"]] += 1
        for key in keys:
            for stat_key, value in values.items():
                buckets[key][stat_key] += value
        for stat_key, value in values.items():
            by_dataset[dataset_key][stat_key] += value

    def finalize(bucket: dict[str, Any]) -> dict[str, Any]:
        count = bucket["num_examples"]
        output = dict(bucket)
        if count:
            output["avg_input_tokens"] = round(bucket["input_tokens"] / count, 2)
            output["avg_support_tokens"] = round(bucket["support_tokens"] / count, 2)
            output["avg_full_record_tokens"] = round(bucket["full_record_tokens"] / count, 2)
        else:
            output["avg_input_tokens"] = 0
            output["avg_support_tokens"] = 0
            output["avg_full_record_tokens"] = 0
        return output

    stats = {
        "name": "general_science_release",
        "version": "v0.1-local",
        "generated_at_utc": datetime.now(timezone.utc).isoformat(),
        "seed": SEED,
        "tokenizer": {
            "name": tokenizer_name,
            "note": "Token counts are computed over question+choices(+image placeholder), support, and serialized full JSON record.",
        },
        "splits": {
            "train": finalize(buckets["split::train"]),
            "test": finalize(buckets["split::test"]),
            "overall": finalize(buckets["overall"]),
        },
        "by_dataset": {key: finalize(value) for key, value in sorted(by_dataset.items())},
        "by_modality": dict(by_modality),
        "source_files": source_stats,
        "deduplication": dedup_stats,
        "sources": SOURCE_LICENSES,
        "outputs": {
            "train": str(TRAIN_PATH.relative_to(RELEASE_DIR)),
            "test": str(TEST_PATH.relative_to(RELEASE_DIR)),
            "stats": str(STATS_PATH.relative_to(RELEASE_DIR)),
            "card": str(CARD_PATH.relative_to(RELEASE_DIR)),
            "readme": str(README_PATH.relative_to(RELEASE_DIR)),
            "scienceqa_images": str(IMAGE_DIR.relative_to(RELEASE_DIR)),
        },
    }
    return stats


def write_jsonl(path: Path, examples: list[dict[str, Any]]) -> None:
    with path.open("w", encoding="utf-8") as handle:
        for example in examples:
            handle.write(json.dumps(example, ensure_ascii=False) + "\n")


def markdown_table(rows: list[list[Any]], headers: list[str]) -> str:
    lines = [
        "| " + " | ".join(headers) + " |",
        "| " + " | ".join("---" for _ in headers) + " |",
    ]
    for row in rows:
        lines.append("| " + " | ".join(str(cell) for cell in row) + " |")
    return "\n".join(lines)


def write_card(stats: dict[str, Any]) -> None:
    source_rows = []
    for dataset, info in SOURCE_LICENSES.items():
        source_rows.append([
            dataset,
            "ARC-Challenge, ARC-Easy" if dataset == "allenai/ai2_arc" else "`null`",
            info["license"],
            info["license_source"],
            info["url"],
        ])

    split_rows = []
    for split_name in ["train", "test", "overall"]:
        split = stats["splits"][split_name]
        split_rows.append([
            split_name,
            split["num_examples"],
            split["image_examples"],
            split["text_examples"],
            split["avg_input_tokens"],
            split["avg_support_tokens"],
            split["avg_full_record_tokens"],
        ])

    dataset_rows = []
    for key, value in stats["by_dataset"].items():
        dataset, subset = key.split("|", 1)
        dataset_rows.append([
            dataset,
            subset,
            value["num_examples"],
            value["image_examples"],
            value["avg_input_tokens"],
            value["avg_support_tokens"],
        ])

    card = f"""# General Science QA Release

## Summary

This release consolidates local science QA snapshots into a unified JSONL format for text and image-text multiple-choice science QA.
It is built from local parquet files only and stops before any public upload.

## Output Files

- `train.jsonl`
- `test.jsonl`
- `stats.json`
- `images/scienceqa/`
- `general_science_card.md`
- `README.md`

## Sources

{markdown_table(source_rows, ["Dataset", "Subset", "License", "License source", "URL"])}

OpenBookQA is present locally but is not included in this main release because its license is not confirmed in the local snapshot. It can be added later as an internal/optional extension after license confirmation.

## Split Construction

- Upstream `train` and `validation` splits are merged into `train.jsonl`.
- Upstream `test` splits are kept as the held-out `test.jsonl`.
- Train/test leakage is checked using normalized question, choices, and answer text.
- If a normalized duplicate appears in held-out test, the matching train example is removed.

## Schema

```json
{{
  "id": "string",
  "dataset": "string",
  "subset": "string or null",
  "split": "train or test",
  "task_type": "multiple_choice_science_qa",
  "modality": "text or image_text",
  "question": "string",
  "image": null,
  "choices": [
    {{"label": "A", "text": "string"}}
  ],
  "answer_label": "string",
  "answer_text": "string",
  "support": "string or null",
  "source_meta": {{}}
}}
```

For ScienceQA image examples, `image` is:

```json
{{"path": "images/scienceqa/<id>.png", "mime_type": "image/png"}}
```

## Counts and Token Statistics

Tokenizer: `{stats["tokenizer"]["name"]}`.

{markdown_table(split_rows, ["Split", "Examples", "Image examples", "Text examples", "Avg input tokens", "Avg support tokens", "Avg full record tokens"])}

## Counts by Dataset

{markdown_table(dataset_rows, ["Dataset", "Subset", "Examples", "Image examples", "Avg input tokens", "Avg support tokens"])}

## Modality Distribution

{markdown_table([[key, value] for key, value in sorted(stats["by_modality"].items())], ["Modality", "Count"])}

## Deduplication

```json
{json.dumps(stats["deduplication"], ensure_ascii=False, indent=2)}
```

## Construction Notes

- SciQ: combines `distractor1`, `distractor2`, `distractor3`, and `correct_answer`, then deterministically shuffles choices.
- AI2 ARC: uses both `ARC-Challenge` and `ARC-Easy`; finds the answer text from `answerKey`, then deterministically shuffles choices.
- ScienceQA: keeps `subject == "natural science"` examples with at least four choices and a valid answer index. Images embedded as parquet bytes are exported to `images/scienceqa/`.
- Choices with more than four options are retained; labels are regenerated from `A`.

## License Notes

This release combines multiple upstream datasets. Downstream redistribution must satisfy all upstream licenses.
ScienceQA license metadata was not present in the local snapshot available to this build; verify the upstream license before public HF/ModelScope upload.

## Recommended Evaluation Input

For model evaluation, use `question`, `image`, and `choices`.
Do not feed `support` unless the task explicitly allows explanations or retrieval context, because `support` may contain answer-revealing information.
"""
    CARD_PATH.write_text(card, encoding="utf-8")
    README_PATH.write_text(card, encoding="utf-8")


def main() -> int:
    RELEASE_DIR.mkdir(parents=True, exist_ok=True)
    IMAGE_DIR.mkdir(parents=True, exist_ok=True)

    raw_items, source_stats = collect_items()
    deduped_items, dedup_stats = deduplicate(raw_items)
    examples = [build_example(item) for item in deduped_items]

    train = [example for example in examples if example["split"] == "train"]
    test = [example for example in examples if example["split"] == "test"]
    write_jsonl(TRAIN_PATH, train)
    write_jsonl(TEST_PATH, test)

    stats = compute_stats(examples, source_stats, dedup_stats)
    STATS_PATH.write_text(json.dumps(stats, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    write_card(stats)

    print(f"Train: wrote {len(train)} examples to {TRAIN_PATH.name}")
    print(f"Test: wrote {len(test)} examples to {TEST_PATH.name}")
    print(f"Total: wrote {len(examples)} examples")
    print(f"Image examples: {stats['splits']['overall']['image_examples']}")
    print(f"Dataset card: {CARD_PATH.name}")
    return 0


if __name__ == "__main__":
    try:
        raise SystemExit(main())
    except Exception as exc:
        print(f"Error: {exc}", file=sys.stderr)
        raise SystemExit(1) from exc