KRISHNAPURI/q-trust-code-corpus-v4

Q-Trust Code Corpus v4

Labelled source-code corpus for cryptographic-code discovery (post-quantum migration). It is the training and evaluation corpus for KRISHNAPURI/q-trust-crypto-detector, and the data behind the in-browser scanner.

• 17,171 file records / 58 source repositories / 8,814 crypto-labelled, 12 languages (C, C++, Java, Go, Rust, Python, JS/TS, Solidity, C#, PHP, Zig, Swift).
• Sources include OpenSSL, BoringSSL, AWS-LC, wolfSSL, Mbed-TLS, GnuTLS, NSS, libsodium, Botan, PQClean, liboqs, Cloudflare CIRCL, RustCrypto, secp256k1, curve25519-dalek, google/tink, pyca/cryptography, and negative controls (numpy, express, svelte, serde, ...).
• Labels are regex-derived (is_crypto = file contains a crypto keyword: RSA/ECC, AES/GCM/ ChaCha20, SHA/BLAKE, X.509, and PQC ML-KEM/Kyber, ML-DSA/Dilithium, SLH-DSA/SPHINCS+, Falcon, HQC). This is a weak label — the attainable detector F1 is bounded by the labeler's own error.

Record schema

{"code": "...", "language": "python", "label": "RSA/ECC", "is_crypto": true,
 "source": "pyca/cryptography", "path": "src/.../rsa.py"}

Split and evaluation protocol

Repo-disjoint, digest-keyed split (SHA-256 of the source name, seed 42): 14,933 train / 2,238 eval records, 0 % train/eval path overlap (verified by an evaluation-leakage audit). The held-out set repeats 197 (source, path) pairs (459 extra records, 22 label conflicts); the primary metric keeps the longest copy per path, chosen without looking at the label, for n = 1,779.

Measured detector results on this corpus (digest split):

View	Gate	Precision	Recall	F1	n
De-duplicated, one per path (primary)	0.50	0.9264	0.9146	0.9205	1,779
Raw records	0.50	0.9201	0.9088	0.9144	2,238
Native split, raw records	0.50	0.9232	0.9238	0.9235	2,566

One-per-path sensitivity across policies: 0.9208 (keep-first), 0.9223 (keep-last), 0.9205 (keep-longest, primary). The full protocol, including dedup policies and recorded negatives, is in EXPERIMENTS.md E-14 → E-16 and in qtrust_ai/artifacts/eval_dedup_v4.json in the GitHub repository.

Related resources

• Detector: KRISHNAPURI/q-trust-crypto-detector
• Live demo: KRISHNAPURI/q-trust-scanner
• Dataset package with the full corpus lineage (v1 → v5, TLS scans, CVEs, CBOMs): KRISHNAPURI/q-trust-datasets
• Source: humoge7502/q-trust

The corpus contains only public source code; no proprietary data. Labels are automated annotations, not human ground truth. License follows the upstream repositories (Apache-2.0 for this packaging; individual source files retain their original licenses).