14Reproducibility
Run these from a clone of the repository on Apple Silicon. Identical requests to the same server return identical answers, so a re-run on the pins below should reproduce these numbers.
make bootstrap # dependencies, pinned weights, readiness check
ember start # warm model server on 127.0.0.1:8765
ember eval run # every item; or --split test, --category <recipe>
ember eval export # this report: Markdown, HTML, figures, raw data
ember eval report --compare results/<a>_results.json results/<b>_results.json| Pin | Value |
|---|---|
| Commit | cadc1ec |
| Dataset SHA-256 | 6a457f6badb7dd53b50a9ce04857d4d01e885d4718f3eb6db9db58658d9f923c |
| Model revision | 17f0b0ad64efb65d273590632833508766b2aae6 |
| Results file | clef-flash_20261003T203810Z_results.json |
| Trace file | clef-flash_20261003T203810Z_trace.jsonl |
| Bootstrap | 1,000 resamples, seed 0, percentile interval |
| Package gut | 0.1.0 |
| Package torch | 2.14.1 |
| Package transformers | 5.18.0 |
| Package mcp | 2.3.0 |