# Bash SFT note 01 evidence

This directory accompanies [`../bash-sft-01.html`](../bash-sft-01.html).
It is the first public packet in a Bash SFT series, separate from the earlier
tab-title namer notes on this site.

## Result

On 500 Mini-trace `task_solution` rows (first 500 of a held-out validation file),
stock `Qwen/Qwen2.5-Coder-1.5B-Instruct` scored exact bash 12.2% and bash F1 41.5%.
The 137-row response-only SFT scored 14.0% and 43.8%. Full-text F1 jumped 41.4% to
66.3% because SFT always emits `Here is a copy-pasteable command`. Device: CUDA
bf16 on a rented NVIDIA RTX 6000 Ada Generation GPU.

This is greedy text vs gold, not executable pass@1. Weights are not in this bundle.
Gold commands and model completions are not published.

## Files

| File | What |
|---|---|
| `index.html` | Packet landing page |
| `summary.json` | Machine score table (stock vs SFT) |
| `report.md` | Same table as markdown |
| `row_metrics.jsonl` | 500 rows of per-example scores only (no commands) |
| `train_summary.json` | 1-epoch train metrics (4.51 s, 13.90 GiB, 137 rows) |
| `mix_manifest.json` | mix counts and holdout families for this note |
| `SHA256SUMS` | SHA-256 of the files above plus this README |

## What is not here

- `model.safetensors` (2.9G, local only)
- Raw Mini session JSONL
- The 137-row train JSONL
- Gold bash and generated completions (dropped after a security pass)