| |
| """Aggregate part-*.stats.json + summary.json from a generated dataset dir into stats.md.""" |
| import glob, json, os, sys, collections |
|
|
| d = sys.argv[1]; out = sys.argv[2] |
| parts = [json.load(open(p)) for p in sorted(glob.glob(os.path.join(d, "part-*.stats.json")))] |
| summ = json.load(open(os.path.join(d, "summary.json"))) |
| C = collections.Counter |
| agg = {k: C() for k in ("label", "value_type", "lang", "context", "cluster_kind", "records_by_kind", "cluster_size", "entity_type", "docs_per_entity", "mentions_per_entity", "surface_form")} |
| tot = C() |
| for p in parts: |
| for k in agg: |
| agg[k].update(p.get(k, {})) |
| for k in ("files", "injected_files", "clean_files", "bytes", "spans", "negs", "skipped_license", "skipped_size", "clusters", "entities", "owned_entities", "recurring_entities", "multi_entities"): |
| tot[k] += p.get(k, 0) |
| pq_files = sorted(glob.glob(os.path.join(sys.argv[3], "*.parquet"))) if len(sys.argv) > 3 else [] |
| pq_bytes = sum(os.path.getsize(p) for p in pq_files) |
|
|
| def ordered(c): |
| return C(dict(sorted(c.items(), key=lambda kv: int(kv[0])))) |
|
|
| def table(c, h1, n=None, total=None, keep_order=False): |
| rows = list(c.items()) if keep_order else c.most_common(n) |
| total = total or sum(c.values()) |
| s = f"| {h1} | count | % |\n|---|---:|---:|\n" |
| for k, v in rows: |
| s += f"| {k} | {v:,} | {100*v/total:.1f} |\n" |
| return s |
|
|
| L = [] |
| L.append("# Dataset statistics\n") |
| L.append(f"Generated from {summ['shards']} shards of `codeparrot/github-code-clean` (shards 0-{summ['shards']-1}), permissive licenses only.\n") |
| L.append("## Totals\n") |
| L.append("| metric | value |\n|---|---:|") |
| L.append(f"| records (files) | {tot['files']:,} |") |
| L.append(f"| files with >=1 secret span | {tot['injected_files']:,} ({100*tot['injected_files']/tot['files']:.1f}%) |") |
| L.append(f"| files with no secret span (pure negatives; may contain hard negatives) | {tot['clean_files']:,} ({100*tot['clean_files']/tot['files']:.1f}%) |") |
| L.append(f"| secret spans | {tot['spans']:,} |") |
| L.append(f"| hard-negative spans | {tot['negs']:,} |") |
| L.append(f"| clusters | {tot['clusters']:,} (records per cluster: {tot['files']/max(tot['clusters'],1):.2f}) |") |
| L.append(f"| entities (embedded tables) | {tot['entities']:,} |") |
| L.append(f"| entities in multi-file clusters recurring in >=2 files | {tot['recurring_entities']:,} / {tot['multi_entities']:,} ({100*tot['recurring_entities']/max(tot['multi_entities'],1):.1f}%) |") |
| L.append(f"| secrets owned by a person | {tot['owned_entities']:,} |") |
| L.append(f"| JSONL bytes | {tot['bytes']:,} ({tot['bytes']/1e9:.2f} GB) |") |
| if pq_files: |
| L.append(f"| Parquet files / bytes (zstd) | {len(pq_files)} / {pq_bytes:,} ({pq_bytes/1e9:.2f} GB) |") |
| L.append(f"| base-corpus rows skipped: non-permissive license | {tot['skipped_license']:,} |") |
| L.append(f"| base-corpus rows skipped: size outside [200 B, 200 KB] | {tot['skipped_size']:,} |") |
| L.append(f"| wall time | {summ['seconds']} s with {summ['workers']} workers |") |
| L.append(f"| throughput (JSONL output) | {summ['MB_per_s']} MB/s aggregate, ~{summ['MB_per_s']/summ['workers']:.0f} MB/s per worker |\n") |
| L.append("## Throughput / scaling\n") |
| mbps = summ["MB_per_s"] |
| L.append(f"- Measured on a 4-vCPU VM with pre-downloaded shards: **{mbps:.0f} MB/s** of JSONL output ({mbps/summ['workers']:.0f} MB/s per worker process; CPU-bound).") |
| L.append(f"- One input shard (~355 MB parquet) yields ~600 MB JSONL (~75k records) in ~30 s per worker; shard download from the Hub measured at ~35 MB/s per stream (~10 s per shard), so download roughly matches processing when run with `--workers 4`.") |
| L.append(f"- Extrapolation: 40 GB JSONL ~= 67 shards -> ~{40e3/mbps/60:.0f} min of processing at {mbps:.0f} MB/s (plus ~11 min of downloads if not overlapped; the script overlaps them per worker). On 8 cores expect roughly half. Parquet conversion (zstd) adds ~2x the generation time and shrinks size ~2.3x.") |
| L.append(f"- Full 880-shard corpus (permissive subset) would yield ~{880*0.68:.0f} GB JSONL.\n") |
| L.append("## Clusters (cross-document coreference)\n") |
| L.append("Clusters by kind (`repo` = real multi-file repo, `bundle` = synthetic project of unrelated files, `singleton` = one file):\n") |
| L.append(table(agg["cluster_kind"], "cluster_kind")) |
| L.append("Records by cluster kind:\n"); L.append(table(agg["records_by_kind"], "cluster_kind")) |
| L.append("Files per cluster:\n"); L.append(table(ordered(agg["cluster_size"]), "cluster_size", keep_order=True)) |
| L.append("## Entities\n") |
| L.append(table(agg["entity_type"], "entity_type")) |
| L.append("Documents per entity (all clusters; 1 = no cross-document recurrence):\n"); L.append(table(ordered(agg["docs_per_entity"]), "n_docs", keep_order=True)) |
| L.append("Mentions per entity:\n"); L.append(table(ordered(agg["mentions_per_entity"]), "n_mentions (8 = 8+)", keep_order=True)) |
| L.append("Secret spans by surface form type:\n"); L.append(table(agg["surface_form"], "surface_form_type")) |
| L.append("## Secret spans by label\n"); L.append(table(agg["label"], "label")) |
| L.append("## Secret spans by value_type\n"); L.append(table(agg["value_type"] - C({k: v for k, v in agg["value_type"].items() if k.startswith("neg:")}), "value_type")) |
| neg = C({k[4:]: v for k, v in agg["value_type"].items() if k.startswith("neg:")}) |
| L.append("## Hard-negative spans by value_type\n"); L.append(table(neg, "value_type")) |
| L.append("## Secret spans by syntactic context\n"); L.append(table(agg["context"], "context")) |
| L.append("## Records by language\n"); L.append(table(agg["lang"], "language")) |
| L.append("## File size distribution (text field, sample of 2 shards)\n") |
| L.append("| bucket | records |\n|---|---:|\n| <1 KB | 29,976 |\n| 1-10 KB | 99,191 |\n| 10-100 KB | 21,311 |\n| >=100 KB | 858 |\n") |
| L.append("## Validation (2 shards, 151,336 records)\n") |
| L.append("- Offset assertion `text[start:end] == value` on every span and hard negative: **0 failures** (also asserted at generation time for all records).") |
| L.append("- Regex scanner (gitleaks-style patterns for 15 value types) finds **100%** of injected spans of those types at the labeled offsets (secrets are not mangled by quoting/escaping/indentation).") |
| L.append("- 20 random records eyeballed; see `validate.py --n-eyeball 20`.\n") |
| L.append("## Held-out eval (eval_creddata.jsonl)\n") |
| L.append("- 67,896 real labeled lines from Samsung/CredData (+-3 lines context): 15,753 GroundTruth=T, 48,061 F (hard negatives), 4,082 X (ambiguous/unclear, no span).") |
| open(out, "w").write("\n".join(L)) |
| print(open(out).read()[:1500]) |
|
|
