convert_openjev.py

import hashlib
import json
from collections import Counter, defaultdict
from datasets import load_dataset
from alea.genbase import emit, q_choice, q_noul, q_score, validate_row, write_jsonl

CONFIGS = {
    "amount-extraction-control-v1": 10000,
    "citation-control-v1": 10000,
    "entity-alignment-control-v1": 10000,
    "email-selection-control-v1": 10000,
    "phone-extraction-control-v1": 10000,
    "context-retention-control-v1": 10000,
    "silent-failure-control-v1": 10000,
    "ir-control-v1": 10000,
    "mailroom-control-v1": 12000,
    "sponsor-segment-control-v1": 8000,
}
MAX_OPTIONS = 16


def parse_json_field(value):
    if not isinstance(value, str):
        return value
    try:
        return json.loads(value)
    except Exception:
        return value


def fingerprint(row):
    raw = json.dumps(row["state"], sort_keys=True, ensure_ascii=False) + json.dumps(row["questions"], sort_keys=True, ensure_ascii=False)
    return hashlib.sha256(raw.encode()).hexdigest()


def convert(ex, config, idx, split):
    state = parse_json_field(ex["state_json"])
    if isinstance(state, str):
        state = state.strip()
    question = ex["question"]
    options = ex["options"]
    target = ex["target"]
    kind = str(ex["kind"]).lower()
    if not isinstance(options, list) or not isinstance(target, list):
        return None
    if len(options) < 2 or len(options) > MAX_OPTIONS or len(options) != len(target):
        return None
    if any(not isinstance(o, (str, dict, list)) for o in options):
        return None
    if kind == "choice":
        q = q_choice("answer", question, options, [float(x) for x in target])
    elif kind == "score":
        q = q_score("score", question, options, [float(x) for x in target])
    elif kind == "noul":
        if len(target) != 2:
            return None
        q = q_noul("answer", question, float(target[1]))
    elif kind in ("multilabel", "multi_label"):
        qs = [q_noul(f"label_{i}",
                     f"Based on the state, {question} Is this criterion true: {opt}?",
                     float(target[i])) for i, opt in enumerate(options)]
        row = emit(f"openjev_{config}", idx, state, qs)
        row["id"] = str(ex.get("id", row["id"]))
        row["provenance"] = {"dataset": "ZefanCai/Open-Jev", "config": config,
                             "group_id": ex.get("group_id"), "split": split,
                             "source": ex.get("source")}
        validate_row(row)
        return row
    else:
        return None
    row = emit(f"openjev_{config}", idx, state, [q])
    row["id"] = str(ex.get("id", row["id"]))
    row["provenance"] = {"dataset": "ZefanCai/Open-Jev", "config": config,
                         "group_id": ex.get("group_id"), "split": split,
                         "source": ex.get("source")}
    validate_row(row)
    return row


def load_split(split):
    rows, seen, groups = [], set(), defaultdict(set)
    stats = Counter()
    next_idx = 0
    for config, limit in CONFIGS.items():
        ds = load_dataset("ZefanCai/Open-Jev", config, split=split, streaming=True)
        cfg_rows = 0
        for ex in ds:
            stats["raw"] += 1
            kind = str(ex.get("kind", "")).lower()
            stats[f"kind:{kind}"] += 1
            row = convert(ex, config, next_idx, split)
            next_idx += 1
            if row is None:
                stats["unsupported_or_invalid"] += 1
                continue
            group = row["provenance"].get("group_id")
            if group is not None:
                groups[config].add(str(group))
            key = fingerprint(row)
            if key in seen:
                stats["duplicate_input"] += 1
                continue
            seen.add(key)
            rows.append(row)
            cfg_rows += 1
            if cfg_rows >= limit:
                break
        stats[f"kept:{config}"] = cfg_rows
    return rows, groups, stats


def main():
    train, train_groups, train_stats = load_split("train")
    validation, val_groups, val_stats = load_split("validation")
    overlap = {c: len(train_groups[c] & val_groups[c]) for c in CONFIGS}
    if any(overlap.values()):
        raise RuntimeError(f"group leakage across train/validation: {overlap}")
    train_keys = {fingerprint(r) for r in train}
    before = len(validation)
    validation = [r for r in validation if fingerprint(r) not in train_keys]
    write_jsonl(train, "data/openjev_train.jsonl")
    write_jsonl(validation, "data/openjev_validation.jsonl")
    report = {
        "dataset": "ZefanCai/Open-Jev",
        "configs": CONFIGS,
        "used_splits": ["train", "validation"],
        "excluded_splits": ["calibration", "test", "ood"],
        "excluded_overlapping_configs": ["release-v2-redistributable", "browser-drone-expansion-v1-redistributable"],
        "train_stats": dict(train_stats),
        "validation_stats": dict(val_stats),
        "groups_shared_by_config": overlap,
        "validation_exact_duplicates_removed": before-len(validation),
        "converted_train_rows": len(train),
        "converted_validation_rows": len(validation),
        "model_inputs": ["state", "question", "options"],
        "privileged_metadata_excluded_from_model_input": ["metadata_json", "record_json", "target", "group_id", "split", "source"],
    }
    json.dump(report, open("data/openjev_conversion_report.json", "w"), indent=2)
    print(json.dumps({"train": len(train), "validation": len(validation),
                      "overlap": overlap, "train_kinds": dict(train_stats),
                      "dev_kinds": dict(val_stats)}, indent=2))


if __name__ == "__main__":
    main()