Datasets:
Submission: forward-metric/sysone-typed-decisions-0.4B (specialist): accuracy 0.8025
Model: https://huggingface.co/forward-metric/sysone-typed-decisions-0.4B
Mode: specialist. Fully fine-tuned on train only, with no other training data. Temperature calibration is fitted on a held-out 10% of train; test was used only for scoring.
Base: jhu-clsp/ettin-encoder-400m (400M encoder), trained with sysone (open source, pip install sysone).
| accuracy | KL from gold | Brier | ECE | latency |
|---|---|---|---|---|
| 0.8025 | KL | 0.0438 | ECE | MS ms per case (5 questions, single request, Colab L4) |
- Scored on all 400 test cases / 2,000 decisions with the formulas in
bench/score_with_jobs.py. Raw predictions are in the model repo (predictions.jsonl) for re-scoring, and the eval-results YAML is on the repo. - At 400M parameters it would be the smallest model at 0.80 or above in the fine-tuned table; the next smallest is 3B. It also uses no data beyond
train. - Answers don't depend on option order: options are read in parallel with shared positions, so the flip rate is 0 by construction.
- Single training run (seed 0), 59 minutes on a Colab L4. More seeds are coming; I'll update this thread with the mean.
To reproduce: the notebook, step 3. To load: SystemOneModel.load("forward-metric/sysone-typed-decisions-0.4B").
Thanks for submitting this.
sysone-typed-decisions-0.4B is now on the card in the fitted / fine-tuned table at #6 with 0.8025 accuracy and Brier 0.0438, tied with Amos 1 large on accuracy.
The original post left KL, ECE and latency as placeholders, so those cells currently show a dash. Please post those three numbers, and the multi-seed mean when ready, and I'll update the row.
Thanks for adding it, and sorry for the placeholders. Here are the three numbers, from the same run and scored with bench/score_with_jobs.py's formulas on all 400 test cases:
| KL from gold | ECE | latency |
|---|---|---|
| 0.0800 | 0.1743 | 40 ms per case (5 questions, single request, p50, Colab L4) |
They're also in results.json in the model repo.
Thanks, the row now has KL 0.080, ECE 0.174 and 40 ms on a Colab L4, matching results.json in the model repo.