TraceTune — Lexsi Track 3 Models & Adapters
Collection
TraceTuned Cohere Tiny Aya (3.35B) LoRA adapters and AgentTune SFT+GRPO policy artifacts for the Lexsi AI Safety & Alignment Hackathon (Track 3). • 1 item • Updated
TraceTune Tiny Aya (3.35B) — AgentTune SFT + GRPO Policy Adapter
TraceTune post-trains Cohere Tiny Aya (CohereLabs/tiny-aya-fire, 3.35B) using Lexsi AgentTune (Two-Tier EventLog Supervised Fine-Tuning + GRPO Reinforcement Learning) and evaluates it with Lexsi AuditKIT (LSAL-1.2 Unified Spine) for autonomous code and security vulnerability triage.
Headline Results (AuditKIT LSAL-1.2 Spine)
| Metric | Untuned Baseline Tiny Aya | TraceTuned Tiny Aya | Improvement Delta |
|---|---|---|---|
| Schema Validity Index (SVI) | 75.0% | 100.0% | +25.0% |
| Loop Trajectory Rate (LTS) | 25.0% | 0.0% | -25.0% (0 Loops) |
| Grounded Citation Precision (GCP) | 0.0% | 100.0% | +100.0% |
| Step Efficiency Ratio (SER) | 1.00 (Incomplete) | 1.00 (Optimal) | 100% Task Completion |
- AuditKIT Provenance Fingerprint:
ak-9e7486b6e67f8e6a - GitHub Repository: https://github.com/grzdev/tracetune
- Dataset & Benchmark Logs: https://huggingface.co/datasets/saulkang/tracetune-trajectories
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support