Spaces:

---

multimodalart

/

jev-decision-index

Like 687

Running

Hi, and thanks again for the re-run of Hopper and for maintaining the Index.

We'd like to request a new, separate row for Hopper (G) 1.2, a general-purpose version of Hopper. The
existing Hopper row (v1.1.1) should stay as it is.

• Weights: HopitAI/hopper-g, revision d60a1d6ca3f3fa25e5daddea16b8ba2b431ebcee (LoRA on Qwen/Qwen3.5-4B, same shape as Hopper).
• Code: https://github.com/hopit-ai/hopper, tag g-1.2.0 (commit 0204f92954f76fd8835f6050c271fc6e75ac3370). The serving code is
  identical to Hopper 1.1.1: same Decider.score entry point, same calibration map, long menus on by default. Only
  the adapter differs, so the run is the same as your Hopper 1.1.1 run with HopitAI/hopper-g in place of
  HopitAI/hopper.
• What we measured on our own run of the 0.2 suite (all 40 benchmarks, A10G, same serving code): balanced raw
  53.50 vs 52.74 for Hopper 1.1.1 (paired bootstrap +0.76, 95 % interval +0.55 to +0.98); balanced skill 38.07 vs
  37.10. These are our numbers, not a claim about the board.
• Licence: research and demo use only, as for Hopper (see the model card).

Happy to answer anything or change anything.