"""Build an immutable train-only expansion while preserving frozen evaluation. Only labeled official TRAIN rows become training or diagnostic decisions. Source validation/test files are projected onto grouping metadata for exclusion only. The existing evaluation files are copied byte-for-byte; no predictions are read. """ import argparse from collections import Counter from datetime import datetime, timezone import hashlib import json import os from pathlib import Path import random import shutil import subprocess import tempfile import unicodedata import urllib.request VERSION = "public-decisions-v2" PROTECTED = ("validation", "calibration", "test", "holdout") SPLITS = ("train", *PROTECTED) SOURCES = { "hellaswag": { "repo": "Rowan/hellaswag", "provider": "huggingface", "revision": "218ec52e09a7e7462a5400043bb9a69a41d06b76", "license": "MIT", "license_reference": "README.md", "expected_train_rows": 39905, "group_columns": ["ctx", "source_id"], "transform": "ctx -> state; four endings -> choices; label -> target; source_id groups", "question": "Which continuation is most plausible given this context?", }, "piqa": { "repo": "ybisk/ybisk.github.io", "provider": "github", "revision": "21edab439af693b961be2f069e8690a88d3b4e37", "license": "AFL-3.0", "license_reference": "piqa/README.md", "expected_train_rows": 16113, "group_columns": ["goal"], "transform": "goal -> state; sol1/sol2 -> choices; train-labels.lst -> target; goal groups", "question": "Which solution best achieves the goal?", }, "commonsenseqa": { "repo": "tau/commonsense_qa", "provider": "huggingface", "revision": "94630fe30dad47192a8546eb75f094926d47e155", "license": "MIT", "license_reference": "README.md", "expected_train_rows": 9741, "group_columns": ["question"], "transform": "question -> state; choices.text -> choices; answerKey mapped through choices.label; question groups", "question": "Which candidate correctly answers the question in the state?", }, } def sha256(path): digest = hashlib.sha256() with Path(path).open("rb") as handle: for chunk in iter(lambda: handle.read(1024 * 1024), b""): digest.update(chunk) return digest.hexdigest() def normalized(text): return " ".join(unicodedata.normalize("NFKC", text).casefold().split()) def fingerprint(text): return hashlib.sha256(normalized(text).encode()).hexdigest() def json_write(path, value): Path(path).write_text(json.dumps(value, indent=2, sort_keys=True) + "\n") def read_rows(path): with Path(path).open() as handle: for line in handle: if line.strip(): yield json.loads(line) def source_keys(family, row): """Return only group/text identities; never examine evaluation labels.""" state = row[{"hellaswag": "ctx", "piqa": "goal", "commonsenseqa": "question"}[family]] if not isinstance(state, str) or not state.strip(): raise ValueError(f"{family}: missing nonempty state") source_group = row.get("source_id") if family == "hellaswag" else state if not isinstance(source_group, str) or not source_group.strip(): raise ValueError(f"{family}: missing source group") return family + ":" + fingerprint(source_group), fingerprint(state) def convert_row(family, index, raw): source = SOURCES[family] group, _ = source_keys(family, raw) if family == "hellaswag": state, choices, target = raw["ctx"], raw["endings"], int(raw["label"]) original_id = str(raw["ind"]) elif family == "piqa": state, choices, target = raw["goal"], [raw["sol1"], raw["sol2"]], int(raw["label"]) original_id = str(index) elif family == "commonsenseqa": state, choices = raw["question"], raw["choices"]["text"] target = raw["choices"]["label"].index(raw["answerKey"]) original_id = str(raw["id"]) else: raise ValueError("Unknown source family") expected_choices = {"hellaswag": 4, "piqa": 2, "commonsenseqa": 5}[family] if (len(choices) != expected_choices or not 0 <= target < len(choices) or any(not isinstance(choice, str) or not choice.strip() for choice in choices) or len({normalized(choice) for choice in choices}) != len(choices)): raise ValueError(f"{family}: invalid choices or target at row {index}") row_id = f"{family}:train:{original_id}" order = list(range(len(choices))) random.Random(fingerprint("expanded-choices:" + row_id)).shuffle(order) return {"id": row_id, "group": group, "family": family, "source_split": "train", "source_record_id": original_id, "source_row_index": index, "source_revision": source["revision"], "state": state, "question": source["question"], "choices": [choices[i] for i in order], "target": order.index(target)} def download_sources(raw_directory): """Download pinned artifacts and project nontraining files to group metadata.""" os.environ.setdefault("HF_HUB_DISABLE_XET", "1") from huggingface_hub import hf_hub_download import pyarrow.parquet as pq raw_directory = Path(raw_directory) raw_directory.mkdir(parents=True, exist_ok=False) train, reserved, provenance = {}, {}, {} for family, source in SOURCES.items(): directory = raw_directory / family directory.mkdir() files = {} def acquire(filename): if source["provider"] == "huggingface": path = Path(hf_hub_download(source["repo"], filename, repo_type="dataset", revision=source["revision"], local_dir=directory)) url = f"https://huggingface.co/datasets/{source['repo']}/resolve/{source['revision']}/{filename}" else: path = directory / filename path.parent.mkdir(parents=True, exist_ok=True) url = f"https://raw.githubusercontent.com/{source['repo']}/{source['revision']}/{filename}" with urllib.request.urlopen(url, timeout=120) as response: with path.open("xb") as handle: shutil.copyfileobj(response, handle) files[filename] = {"sha256": sha256(path), "bytes": path.stat().st_size, "url": url, "local_path": str(path.relative_to(raw_directory.parent))} return path acquire(source["license_reference"]) reserved[family] = [] split_counts = {} if source["provider"] == "huggingface": for split in ("train", "validation", "test"): path = acquire(f"data/{split}-00000-of-00001.parquet") columns = None if split == "train" else source["group_columns"] rows = pq.read_table(path, columns=columns).to_pylist() split_counts[split] = len(rows) if split == "train": train[family] = rows else: reserved[family].extend(rows) else: rows = list(read_rows(acquire("piqa/data/train.jsonl"))) labels = acquire("piqa/data/train-labels.lst").read_text().splitlines() if len(rows) != len(labels): raise ValueError("PIQA train/label count mismatch") train[family] = [{**row, "label": int(label)} for row, label in zip(rows, labels)] split_counts["train"] = len(rows) for split, name in (("validation", "valid.jsonl"), ("test", "tests.jsonl")): # Only the goal participates in exclusions; no evaluation label file is fetched. rows = [{"goal": row["goal"]} for row in read_rows(acquire("piqa/data/" + name))] split_counts[split] = len(rows) reserved[family].extend(rows) if len(train[family]) != source["expected_train_rows"]: raise ValueError(f"{family}: pinned train row count changed") provenance[family] = {**source, "files": files, "source_counts": split_counts, "evaluation_access": "group/text metadata only; no evaluation labels used"} print(json.dumps({"event": "source_ready", "family": family, "counts": split_counts}), flush=True) return train, reserved, provenance def prepare_dataset(base, output, source_train, source_reserved, source_provenance, max_new_per_family=16000, diagnostic_per_family=128, seed=917): """Write a complete new dataset. All source rows passed here are in-memory data.""" base, output = Path(base).resolve(), Path(output).resolve() if output == base or output.is_relative_to(base): raise ValueError("Expanded output must be separate from its immutable base") if max_new_per_family < 0 or diagnostic_per_family < 0: raise ValueError("Sampling limits must be nonnegative") base_manifest = json.loads((base / "manifest.json").read_text()) if set(base_manifest["split_sha256"]) != set(SPLITS): raise ValueError("Base dataset must have exactly the five frozen decision splits") for split, checksum in base_manifest["split_sha256"].items(): if sha256(base / f"{split}.jsonl") != checksum: raise ValueError("Base dataset hash mismatch: " + split) original = list(read_rows(base / "train.jsonl")) if any(row["family"] in {"social", "socialiqa", "social_i_qa"} for row in original): raise ValueError("Social IQA must never enter training") original_ids = {row["id"] for row in original} if len(original_ids) != len(original): raise ValueError("Duplicate original training ID") old_train_text = {fingerprint(row["state"]) for row in original} reserved_groups, reserved_text = set(), set() for split in PROTECTED: for row in read_rows(base / f"{split}.jsonl"): # Label/choice content does not affect exclusions or sampling. reserved_groups.add(row["group"]) reserved_text.add(fingerprint(row["state"])) if {row["group"] for row in original} & reserved_groups: raise ValueError("Base dataset has train/evaluation group overlap") additions, diagnostics, audit = [], [], {} seen_ids, seen_text = set(original_ids), set(old_train_text) for family in SOURCES: official_groups, official_text = set(), set() for row in source_reserved[family]: group, text = source_keys(family, row) official_groups.add(group) official_text.add(text) removed, candidates = Counter(), [] for index, raw in enumerate(source_train[family]): try: row = convert_row(family, index, raw) except (ValueError, TypeError, KeyError, IndexError): removed["invalid_source_row"] += 1 continue text = fingerprint(row["state"]) if row["group"] in official_groups or text in official_text: removed["official_evaluation_overlap"] += 1 elif row["group"] in reserved_groups or text in reserved_text: removed["original_evaluation_overlap"] += 1 elif row["id"] in seen_ids or text in seen_text: removed["duplicate_training_identity_or_state"] += 1 else: seen_ids.add(row["id"]) seen_text.add(text) candidates.append(row) groups = sorted({row["group"] for row in candidates}, key=lambda group: fingerprint(f"diagnostics:{seed}:{group}")) if len(groups) <= diagnostic_per_family: raise ValueError(f"{family}: insufficient groups after exclusions") diagnostic_groups = set(groups[:diagnostic_per_family]) diagnostic_rows = {} train_rows = [] for row in candidates: if row["group"] in diagnostic_groups: diagnostic_rows.setdefault(row["group"], row) else: train_rows.append(row) # Sampling is deterministic and independent of input/parquet row ordering. train_rows.sort(key=lambda row: (fingerprint(f"train:{seed}:{row['id']}"), row["id"])) selected = train_rows[:max_new_per_family] if max_new_per_family else train_rows additions.extend(selected) diagnostics.extend(diagnostic_rows[group] for group in groups[:diagnostic_per_family]) audit[family] = {"raw_train_rows": len(source_train[family]), "removed": dict(removed), "official_reserved_groups": len(official_groups), "diagnostic_groups": len(diagnostic_groups), "diagnostic_group_rows_excluded_from_training": len(candidates) - len(train_rows), "available_training_rows": len(train_rows), "retained_training_rows": len(selected), "cap_excluded_rows": len(train_rows) - len(selected)} if {row["group"] for row in additions} & {row["group"] for row in diagnostics}: raise ValueError("Expanded train/diagnostic group overlap") if {fingerprint(row["state"]) for row in additions} & reserved_text: raise ValueError("Expanded training state overlaps frozen evaluation") output.mkdir(parents=True, exist_ok=False) for split in PROTECTED: shutil.copyfile(base / f"{split}.jsonl", output / f"{split}.jsonl") original_bytes = (base / "train.jsonl").read_bytes() if original_bytes and not original_bytes.endswith(b"\n"): raise ValueError("Original training JSONL must end in a newline for byte-exact replay") # Original replay is a byte-exact prefix. The trainer shuffles all rows per epoch. with (output / "train.jsonl").open("wb") as handle: handle.write(original_bytes) for row in additions: handle.write((json.dumps(row, sort_keys=True) + "\n").encode()) diagnostic_path = output / "diagnostics" / "new_sources.jsonl" diagnostic_path.parent.mkdir() diagnostic_path.write_text("".join(json.dumps(row, sort_keys=True) + "\n" for row in diagnostics)) split_hashes = {split: sha256(output / f"{split}.jsonl") for split in SPLITS} protected_hashes = {split: base_manifest["split_sha256"][split] for split in PROTECTED} if any(split_hashes[split] != expected for split, expected in protected_hashes.items()): raise ValueError("Frozen evaluation changed during copy") counts = {**base_manifest["split_family_counts"], "train": dict(Counter(row["family"] for row in [*original, *additions]))} manifest = {"version": VERSION, "created_utc": datetime.now(timezone.utc).isoformat(), "base_dataset": {"path": str(base), "manifest_sha256": sha256(base / "manifest.json"), "split_sha256": base_manifest["split_sha256"]}, "sources": {**base_manifest["sources"], **source_provenance}, "source_script_sha256": sha256(__file__), "split_sha256": split_hashes, "protected_split_sha256": protected_hashes, "split_family_counts": counts, "diagnostics": {"path": str(diagnostic_path.relative_to(output)), "sha256": sha256(diagnostic_path), "family_counts": dict(Counter(row["family"] for row in diagnostics)), "selection_eligible": False, "source_split": "train"}, "sampling": {"seed": seed, "max_new_per_family": max_new_per_family, "diagnostic_groups_per_family": diagnostic_per_family, "trainer_sampler": "unchanged uniform rows without replacement per epoch", "original_replay_rows": len(original), "original_replay_bytes": len(original_bytes), "original_replay_sha256": base_manifest["split_sha256"]["train"], "new_training_rows": len(additions)}, "audit": audit, "group_leakage": False, "holdout": "Original Social IQA holdout preserved byte-for-byte; no Social IQA training", "normalization": "Unicode NFKC, casefold, collapsed whitespace, SHA256", "limitations": "Exact normalized states/groups only; semantic duplicates and pretraining contamination not excluded. New-source diagnostics are outside fixed checkpoint selection."} json_write(output / "manifest.json", manifest) return manifest def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--base", required=True) parser.add_argument("--output", required=True) parser.add_argument("--max-new-per-family", type=int, default=16000) parser.add_argument("--diagnostic-per-family", type=int, default=128) parser.add_argument("--seed", type=int, default=917) args = parser.parse_args() output = Path(args.output).expanduser().resolve() repository = Path(__file__).resolve().parents[1] if output.exists() or output.is_relative_to(repository): parser.error("Output must be a new directory outside the source repository") output.parent.mkdir(parents=True, exist_ok=True) staging = Path(tempfile.mkdtemp(prefix=f".{output.name}-preparing-", dir=output.parent)) train, reserved, provenance = download_sources(staging / "raw") prepared = staging / "dataset" manifest = prepare_dataset(args.base, prepared, train, reserved, provenance, args.max_new_per_family, args.diagnostic_per_family, args.seed) shutil.move(staging / "raw", prepared / "raw") manifest["source_git_commit"] = subprocess.check_output(["git", "rev-parse", "HEAD"], cwd=repository, text=True).strip() manifest["source_git_status"] = subprocess.check_output(["git", "status", "--porcelain"], cwd=repository, text=True) json_write(prepared / "manifest.json", manifest) if output.exists(): raise ValueError("Output appeared during preparation; refusing overwrite") prepared.rename(output) staging.rmdir() print(json.dumps({"event": "expanded_dataset_ready", "path": str(output), "manifest_sha256": sha256(output / "manifest.json"), "counts": manifest["split_family_counts"], "audit": manifest["audit"]}, indent=2), flush=True) if __name__ == "__main__": main()