# Bash SFT note 03

Status: research note, not a paper
Date: 2026-08-30

## Headline

- Mix composition Arm A (full ~75/15/10 task/repair/trajectory) has the lowest validation loss: 0.7180.
- Kept product checkpoint: last40 wd=0.1, GFR 2,402 rows, 1e-5 cosine, weight decay 0.1: 92/98 locked macOS file-search, exact 86/98.
- Harvest mix continuation: 91/98. Hardfam 3x upsample: 91/98. cosine_with_restarts: 80/98. Kimi K3 hardfam continue: 92/98, not kept.
- Matched-prompt comparison: kept checkpoint 90/98, exact stock base 8/98, and NL2Shell 0.8B specialist 4/98.

## Recommendation

Test the 92/98 checkpoint on real file-search first. Do not upsample the remaining six miss families again. Do not publish training JSONL.

## Claim boundary

92/98 is the frozen product-prompt result on a within-generator development set. The kept checkpoint scored 90/98 in the separate matched-prompt baseline run. Handoff canary remains 0/1. Mix-A val loss is not GFR pass@1. Training corpora are unpublished.