Datasets:

---

LocalLLaMA

/

typed-decisions

Like 213

Follow

LocalLLaMA 234

Hi! Submitting linne-oak 1.0 for the zero-shot table.

• Model: PNPM244/linne-oak-1-0 (31B, Gemma-4-31B-it fine-tune, Apache-2.0, open weights), revision 6db14b4
• Mode: general, zero-shot. No train split, no test rows and nothing derived from them was used for training, fitting or calibration. Probabilities are the model's raw output, no temperature fitting.
• Request shape: one POST /v1/systemone per case with the state and all five questions exactly as in the dataset (400 requests, 2,000 decisions, 0 errors). Served with SGLang, full-attention layers bidirectional, as described in the model card.

By question type: noul 0.848, choice 0.708, score 0.680.
By workflow: agent_trace_observability 0.656, customer_service 0.802, invoice_processing 0.774, security_incidents 0.724.

Accuracy is argmax against the gold label; KL is KL(gold || model); Brier is the per-decision squared error against the gold distribution. My scorer reproduces your Uniform reference row (KL 0.444, Brier 0.238). I left ECE out because definitions differ between submitters.

Latency: p50 1.3 s per case (five questions) with 8 concurrent requests on a single GPU; self-reported and not comparable with the hosted rows.

The KL is high because the model puts most of its mass on one option, similar to the Jev row. Happy to share the raw per-case outputs if useful.