A LoRA fine-tune of WeiboAI/VibeThinker-3B that triages bug-bounty / vulnerability-disclosure submissions into a structured verdict — disposition, severity, confidence, and a rationale — and is hardened against prompt-injection and AI-generated «slop» reports. > Project name: VibeBounty. This repo hosts the trained LoRA adapter (mlx-lm > format); fuse it onto the base model to get a standalone model. Given a report (title, asset, description, steps, impact), it emits a JSON verdict over a 9-class disposition taxonomy: validimpactful · validlow · corroboratedsurge · likelyduplicate · outofscope · theoreticalnopoc · selfinflicted · acceptedrisk · slop plus a severity estimate, a confidence gated by claim-reliability, and questions for the researcher. Or load the base + adapter directly with mlx-lm without fusing (—adapter-path vibebounty/adapters). — Base: WeiboAI/VibeThinker-3B (Qwen2.5-3B lineage) — Method: LoRA (rank 16, scale 20, all 36 layers; q/k/v/o + MLP proj), maskprompt — Iters: 2000, batch 8, seq 2048, lr 1e-4, AdamW — Data: ~18k bug-bounty reports labeled from real disclosure outcomes (substate / severity / bounty / CVE), rendered as chat with reasoning targets — Train…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: macmacmacmac
Теги: mlx, lora, security, bug-bounty, vulnerability-triage, vibethinker, llm-as-judge, en
Лайков: 3 | Загрузок: 0
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
