Spaces:

---

multimodalart

/

jev-decision-index

Like 831

Running

We scored a new AJev adapter ourselves with the kit and opened a PR: https://github.com/apolinario/decision-index/pull/215

• AJev lm6q: andyzhang232/ajev-gemma4-31b-lm6q, a LoRA (r 64) on google/gemma-4-31B-it. 65.08 public index (0.3), 73.36 raw. Run, scores and harness: andyzhang232/ajev-gemma4-31b-lm6q-decision-index
• complete: true, all 140,620 requests ok. Nothing truncated, no options removed. bf16, Apache-2.0. Probabilities come from the option-label token logits in one forward pass (temperature 1). No output tokens, no thinking, no reread.
• Code: cmzy/ajev-infer 0.7.0. A Jev /v1/systemone server on vLLM: python -m ajev.serve_vllm --base-model google/gemma-4-31B-it --adapter andyzhang232/ajev-gemma4-31b-lm6q.
• Single-request latency on one RTX PRO 6000 (vLLM, bf16, 506-request public sample): median 139 ms, mean 620 ms, p80 540 ms.

Not affiliated with TypeSafe AI.