K-Decision-9B v2 (ONYULTEA) — 0.8295 accuracy (fitted)

#78
by ONYULTEA - opened

K-Decision-9B v2 — 0.8295 accuracy (test split, 400 cases / 2,000 decisions)

Category: fitted (fine-tuned on the public train split, per benchmark rules)

Results (this repo's test split, all 400 cases)

Metric Value
Accuracy 0.8295
KL from gold 0.6152
Brier 0.1085
ECE (15-bin) 0.1651
Evaluated decisions 2,000
Errors / unsupported 0 / 0
Reported latency ~298 ms median (int8, single 24GB GPU)

KL/Brier computed on 1,200/2,000 decisions where option key mapping was recoverable (all noul questions + matched choice questions).

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.

This model (v2) was trained with identical hyperparameters to v1 (0.818) but a different random seed. GPU non-determinism in int8 quantized training produced a different convergence path.

Reproducing

git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B-v2 --port 8111
# then evaluate the 400 test cases against http://127.0.0.1:8111/v1/systemone
LocalLLaMA org

Added K-Decision-9B-v2 to the fitted table at rank 1 of 52, on your reported 0.8295. It sits just above ZhiPan-Z3-Pro (0.8265). PR: https://huggingface.co/datasets/LocalLLaMA/typed-decisions/discussions/79

No predictions file is public, so the row could not be rescored. It is marked as creator-reported, and the Kind cell notes that it is a second training seed of K-Decision-9B with the same recipe. v1 stays at rank 4.

KL, Brier and ECE are blank, as they are for v1. Your KL and Brier cover 1,200 of the 2,000 decisions, and your ECE uses 15 bins where the card uses 10. If you publish the full test predictions (all 400 cases, with every option's probability), the row can be rescored with bench/score_with_jobs.py and those columns filled in.

Full test predictions are now public: evaluation/predictions.jsonl

All 400 cases × all questions × every option's probability, mapped to the original criteria keys from the test split. This should be scorable with to fill in the KL, Brier, and ECE columns.

Sign up or log in to comment