Datasets:
Submission: skamalj/signpost-d-4b (4.5B, general, zero-shot): 0.712 acc / 0.1615 KL / 0.0853 Brier / 0.0914 ECE
Model: skamalj/signpost-d-4b, Qwen3.5-4B, a LoRA soup of three seeds, merged fp16 (Apache-2.0).
Mode: general, zero-shot. Trained only on our own generated data in invented domains: no Typed Decisions rows, workflows or question schemas. The train split was analysed to brief our data generation and to compare candidates, never trained or fitted on (details on the card). Four test runs for this model and its candidates. The final model was chosen on train and scored once; the published weights were then re-scored (identical) and scored on HF Jobs.
Setup. test split, 400 cases / 2,000 decisions, one request per case, all answered. Scored on HF Jobs (l40sx1) with bench/score_with_jobs.py and the README serve flags.
| Accuracy | KL from gold | Brier | ECE | p50 per case |
|---|---|---|---|---|
| 0.712 [0.691, 0.734] | 0.1615 | 0.0853 | 0.0914 | 62 ms (Colab G4, RTX PRO 6000) |
By type: noul 0.788 (600), choice 0.692 (600), score 0.670 (800). The 95% interval is a case bootstrap.
- Predictions:
typed_decisions_test_predictions.jsonon the model repo. - Eval results:
.eval_results/skamalj__signpost-d-4b.yaml(merged; notes state the mode and data).
A note on bench/score_with_jobs.sh: on the README's image (pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel) it stops at pip install … --torch-backend=auto, because --torch-backend is a uv pip option. We ran the same steps with uv:
bash -c 'cd /tmp; pip install -q uv datasets requests; uv pip install --system -q -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly; python /repo/bench/score_with_jobs.py $MODEL_ID; python /repo/bench/submit_results.py $MODEL_ID results.json --mode general'
