KRPur/q-trust-codebert

Q-Trust CryptoCodeDetector (CodeBERTa fine-tune)

Fine-tuned crypto-usage discovery model from Q-Trust (qtrust_ai/ intelligence layer). Detects cryptographic API usage and algorithm families in source code — the discovery stage that feeds CBOM generation and PQC migration planning.

Training

• Corpus: 13,973 real code files — SolidiFI, SmartBugs, EIPs, WebAuthn blockchain contracts, OSS crypto repos
• Schedule: 4-epoch GPU fine-tune (A100), deterministic seed (same seed → same F1)
• Dataset: KRPur/q-trust-datasets (code_corpus.json)

Held-out results (repo-disjoint, n=2415)

Metric	Q-Trust ensemble	Rules-only	Majority	Random
F1	0.9525	0.673	0.8683	0.5981
Precision	0.952	0.979	—	—
Recall	0.953	0.513	1.0	—

Source: qtrust_ai/artifacts/benchmark_comparison.json (seed 42) in the GitHub repo.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
m = AutoModelForSequenceClassification.from_pretrained("KRPur/q-trust-codebert")
t = AutoTokenizer.from_pretrained("KRPur/q-trust-codebert")