Datasets:
K-Decision-9B (ONYULTEA) — 0.818 accuracy (fitted)
K-Decision-9B — 0.818 accuracy (test split, 400 cases / 2,000 decisions)
Category: fitted (fine-tuned on the public train split, per benchmark rules)
- Model: ONYULTEA/K-Decision-9B (Apache-2.0)
- Base: Qwen3.5-9B + int8 LoRA (r=32), trained on a single 24GB consumer GPU
- Method: single-pass typed decisions — a JSON skeleton with per-field markers is appended to the prompt; the logit immediately before each marker predicts that field's answer symbol, constrained-softmax over legal options only. All fields answered in one forward pass, no decoding loop.
- Calibration: per-question-type temperature fitted on held-out data; full option probability distributions are returned for every choice.
Results (this repo's test split, all 400 cases)
| Metric | Value |
|---|---|
| Accuracy | 0.818 |
| Evaluated decisions | 2,000 |
| Errors / unsupported | 0 / 0 |
Reproducing
git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B --port 8111
# then evaluate the 400 test cases against http://127.0.0.1:8111/v1/systemone
The served model implements the System One request/response format (state + typed questions -> probabilities), compatible with this benchmark's per-question gold structure.
Supplementary metrics (requested table format)
| Metric | Value |
|---|---|
| Accuracy | 0.818 (initial run) / 0.8195 (re-run) |
| KL from gold | 0.358* |
| Brier | 0.052* |
| Reported latency | ~300 ms median (single-request, int8 on one 24GB GPU) |
*KL/Brier are computed over the directly matchable subset (all noul questions; choice questions where option order could be recovered). Full-protocol numbers available on request — happy to rerun with the official scorer if maintainers share the scoring script interface.
Added to the fitted table at rank 2 on your reported 0.818. KL, Brier and ECE are blank for now because your numbers covered only part of the decisions. The scorer is bench/score_with_jobs.py in this repo. If you run it on all 400 test cases and post the results here, those columns can be filled in.