Spaces:

---

multimodalart

/

jev-decision-index

Like 720

Running

Hi @multimodalart , could you add Image Hopper to the vision index? It is a separate entry from Hopper (text): Hopper's
official path takes no images, Image Hopper's does.

• Weights: https://huggingface.co/HopitAI/image-hopper @ 0048827197777c6e495872aa9f2a55f6a608ebfe (Apache-2.0).
  LoRA adapter on Qwen/Qwen3.5-9B @ c202236235762e1c871ad0ccb60c8ee5ba337b9a (base not redistributed).
• Source: https://github.com/hopit-ai/image-hopper @ 9d1a8d91e2c41ad41d97c65a8aaf346ef38c4d4e.
• Readout: one forward pass per decision (one prefill, no generation); probabilities from the option-letter logits,
  FP32 softmax, temperature per route × answer type (calibration.json).
• Routing: a deterministic text rule (router-rules/v1, question and option text only) sends screen/geometry
  questions to the LoRA and everything else to the frozen base. Images on the default route are capped at 768 image
  tokens each; screen/geometry images stay at native resolution.
• Contract: serves /v1/systemone (choice with 2 to 26 options, noul); up to 4 images per request; score
  questions are not supported. The Dockerfile in the repository starts the server exactly as we run it (one GPU, bf16,
  batch 1, about 20 GB peak). It should work with the kit's http engine.
• Data: the adapter was trained on our own generated data plus GroundCUA, VisA and DocLayNet (training split);
  none of the vision index's benchmark datasets were used for training or calibration. Disclosure: our internal
  development and evaluation sets included items from CharXiv, CORD and Mind2Web, which informed which version we
  kept, but no weights or temperatures were fitted on them.

Thanks for running the index!