aya-expanse-8b-lexsi-v6 — exploratory spend-governance LoRA (NOT the submission adapter)
LoRA adapter continuing training from
Papajams/aya-expanse-8b-lexsi-v4
on 600 additional targeted arithmetic SFT examples (sft_arith.jsonl in the
dataset), for the Lexsi AI Safety & Alignment Hackathon (Track 3).
Published for transparency — read before use
On the canonical 70-scenario held-out suite (bf16), v6 regressed the primary safety metric versus v4:
| metric | base | v4 (submission at the time; superseded by v7) | v6 (this adapter) |
|---|---|---|---|
| accuracy | 58.6% | 78.6% | 80.0% |
| unsafe enforced | 17/48 | 2/48 | 5/48 |
| contract valid | 70/70 | 70/70 | 70/70 |
The arithmetic-weighted continuation made the model more permissive: it fixed
several arithmetic/clarification cases but newly unsafe-allowed judgment-heavy
scenarios (mid-trajectory scope creep, recurring commitments). We publish the
weights so the negative result is reproducible, not because we recommend this
checkpoint. Use aya-expanse-8b-lexsi-v4 for the governed-spend task.
Per-scenario outputs and metrics: results70/ in
Papajams/lexsi-governance.