license: mit
base_model: Qwen/Qwen3-4B-Instruct
tags:
- tracea
- dynamic-analysis
- reverse-engineering
- agents
- runtime-traces
- lora
- peft
library_name: peft
pipeline_tag: text-generation
tracea-explainer β trace β plain-English behavior reports (planned)
π§ Status: adapter training scheduled for Week 4 (target: Nov 10, 2026). This card will be updated with metrics, samples, and the adapter download when training completes on TraceBench.
A LoRA adapter for Qwen3-4B-Instruct that reads Evidence Record v1 streams produced by tracea (hooks, syscall summaries, HTTP/TLS/DNS evidence) and writes a grounded plain-English behavior report β every claim tied to an evidence record.
Why
Agents trace software, but raw evidence is verbose and easy to misread. tracea-explainer is the "narrator": it turns structured runtime evidence into human/agent-readable reports without inventing claims β trained to abstain when evidence doesn't support a statement (the "limitations" discipline of Evidence Records).
Training plan
| Item | Value |
|---|---|
| Base | Qwen/Qwen3-4B-Instruct |
| Data | TraceBench train split (trace JSONL β narrative) |
| Method | LoRA (r=16, Ξ±=32), bf16, seq len 8k |
| Eval | held-out scenarios; factuality vs. ground truth; abstention rate |
| Serving | vLLM/llama.cpp; wired into the tracea-demo Space "Auto-explain" tab |
Intended use
- Narrating authorized dynamic-analysis sessions (your apps, devices, captures)
- Security education, debug triage summaries, agent report generation
Out of scope
Closed-source targets without authorization; malware development; DRM circumvention.
Links
- π Code: Hisernberg/tracea
- π Data: Nabidnur/tracea-bench
- π Demo: Nabidnur/tracea-demo