review-v3d-clean-data

The data that trained and evaluated reinaldosaraiva/review-v3d-clean: single diff hunks from nine open-source projects (C, Python, Ruby), each labelled with a JSON list of concrete defects or an empty list.

Files

File	Rows	Content
data/train.jsonl	2681	chat-format training rows: 749 teacher-labelled hunks + 1932 seeded contrastive rows (boolean family replicated, ids with ~rN)
data/valid.jsonl	254	validation rows, same composition, no replicas
data/gold.jsonl	530	evaluation gold: 319 clean and 211 defective hunks (by label)
data/gate_pairs.jsonl	20	seeded bug / un-mutated twin pairs used by the A3 and paired gates
data/gate_test.jsonl	20	ids of the gate pairs (no code; provenance is in gate_pairs.jsonl)
NOTICE	—	per-repository attribution and licenses

Every row that contains code carries repo, source_url, commit, path and license. Chat rows also carry kind: teacher-label (label from Qwen/Qwen3-Coder-30B-A3B-Instruct, then filtered and capped at 8.4 % positives) or sown-pair (a deterministic mutation such as &&→||, <→<=, removed !, with the finding written by the generator; the un-mutated twin is labelled {"findings":[]}).

Gold set

• gold_by: manual — Django security-fix hunks (diff reversed) and clean hunks from the same commits; an AI agent drafted each label on a review sheet and the owner approved the sheet in one batch. sown — seeded defects with a known line, from the mutation generator. prelabel-model — clean hunks pre-labelled by an unrecorded model in an earlier round, confirmed clean by an AI agent and approved by the owner in one batch. harness-review — clean hunks from commits without a fix, pre-labelled by Qwen3.8-Flash-Next, read by an AI agent harness and kept only when it saw no plausible defect. At dataset construction, no row was confirmed by a person reading it hunk by hunk, and no gold label comes from the teacher that labelled the training data. A later 100-item human audit was completed on 2026-10-14: agreement with the original labels was 88/100 (Wilson 95%: 80.19–93.00%), with Cohen's kappa 0.67123. Human authorship was declared by the owner; there was no second independent human review. Twelve disagreements remain without final human adjudication and the dataset labels are unchanged. This is a stratified sample; the unweighted agreement is not a dataset-wide accuracy estimate. See the published audit record.
• A separate AI-assisted technical follow-up proposes quarantine of four textual logic mutants and two label changes, retains the test-coverage category, and records scoped counterevidence for the other disputes. Windows performance and full Rails/database integration remain unverified locally. The accompanying metric scenarios are post hoc sensitivity analyses on existing predictions, not improved model performance or a replacement for this frozen dataset.
• Split: held-out = int(sha256(id), 16) % 2 == 1 (276 rows); the rest (254) is dev, the only part used for checkpoint selection.
• One known mislabelled clean row from the earlier gold (b1bb25079e285f11, a real bug fixed upstream in requests commit 83eff32e) was removed.

Sources

Django, Flask, requests, FastAPI, Linux, FRR, Open vSwitch, OVN and Rails at pinned snapshots (see NOTICE). Vendored third-party directories were excluded, four hunks flagged by gitleaks were excluded, and every row was checked against a private corpus (by id, commit and whitespace-normalized diff) so that no private code is included.

License

Labels and annotations: CC BY 4.0. The code inside each row keeps the license of its origin, recorded per row — this includes GPL-2.0 code (Linux, FRR, the OVS datapath), BSD-3-Clause, MIT and Apache-2.0. If you redistribute rows, follow the license in the row's license field.