Datasets:

LocalLLaMA
/

typed-decisions

e12ex2/jev-vision-27b, a LoRA on Qwen3.8-27B. It never saw this benchmark's train split or these workflows. Scored on test: 400 cases, one POST /v1/systemone per case with all 5 questions, full distributions.

Accuracy	KL from gold	Brier	ECE	p50 latency
0.714	0.318	0.129	0.026	199 ms

By type: noul 0.812, choice 0.697, score 0.654.
By workflow: customer_service 0.772, invoice_processing 0.732, security_incidents 0.724, agent_trace_observability 0.628.
Latency: one client, self-hosted on an H100 (vLLM, bf16), so not comparable with the hosted rows.
Metric definitions: Brier is the multiclass sum of squared errors per decision, averaged over the 2,000 decisions. ECE uses 10 bins on top-option confidence. score questions are answered as a distribution over the rubric levels plus the expected level.

Predictions are available on request if you'd like to re-score with your own scorer.

codelion

LocalLLaMA org 4 days ago

Added to the zero-shot table at #5: 0.714 accuracy, KL 0.318, Brier 0.129, ECE 0.026. It's marked as self-reported, with the 199 ms latency noted as self-hosted on an H100, since that isn't comparable with the hosted rows. Thanks for the per-type and per-workflow breakdown.

e12ex2

3 days ago

no problem

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment