Image Hopper

Image Hopper answers multiple-choice and yes/no questions about images with a calibrated probability for every option, in one forward pass and without generating text. It is built on Qwen/Qwen3.5-9B (revision c202236235762e1c871ad0ccb60c8ee5ba337b9a) and is designed for the Image JevBench /v1/systemone contract.

How it works

• One base, two routes. A fixed text rule (router-rules/v1) reads only the question and option text. Questions about on-screen elements (markers, buttons, menus, windows, …) or geometry (angles, triangles, circles, lengths, …) use a LoRA adapter; every other question uses the unmodified base model. Both routes share one loaded model and one forward pass.
• Readout. Options are lettered A, B, C, …; the answer distribution is the FP32 softmax of the next-token logits for those letters at the last prompt position.
• Calibration. Logits are divided by a temperature chosen by route and answer type (fitted by maximum likelihood with shrinkage toward a global temperature). Calibration never changes which option ranks first.
• Images use the base processor's native resolution route; on the default route each image is capped at 768 image tokens (screen/geometry questions keep full resolution); up to four images per request.

Training and calibration data

The adapter (LoRA r64, one epoch, about 5,900 examples) was trained on real images with their datasets' own labels — desktop screenshots from GroundCUA (training applications only), document pages from DocLayNet, industrial-defect photos from VisA and openly licensed photo collections — plus original procedural renders (charts, documents, tables, geometry, interface mock-ups). No benchmark items were used for training or calibration, and every training image was checked against the benchmark's reference images for near-duplicates.

The default-route temperatures were fitted on a held-out calibration split of our own data; the screen/geometry temperatures on held-out real screenshots and geometry diagrams that were never used for training.

Serving

The Dockerfile starts the server with --serving-options lean_lora,uint8_pixels,device_readout,image_token_cap_default=768. The first three options change only how the computation runs, not the outputs; the image cap on the default route reduces cost per decision.

Usage

pip install -r requirements.txt
python -m open_decisions.image_jev.release.server --offline --adapter adapter --port 8080 --serving-options "lean_lora,uint8_pixels,device_readout,image_token_cap_default=768"

POST /v1/systemone:

{"images": ["data:image/png;base64,..."],
 "questions": {"decision": {"type": "choice", "instructions": "Which shelf is empty?",
   "criteria": {"top": "Top shelf", "bottom": "Bottom shelf"}}}}

The response has model, token usage and answers; a choice answer holds the selected option and a probability per option, a noul answer holds P(yes).

Evaluation

• Image JevBench v0.3.0 (6 Oct 2026): composite 15.3, #16 of 44; Capability 51.7 (Intelligence 32, Calibration 71); measured cost $0.077 per 1,000 decisions, above the Jev-class cost limit. v0.3.0's sealed draw includes rating-scale (score) questions, which this release does not answer (see limitations).
• Jev Decision Index, vision track: scoring requested (discussion #50); results will be added when published.

Intended use and limitations

For research and evaluation of image-grounded decisions where callers supply the options. Answers choice and noul questions only; score (rating-scale) questions are rejected. Not a safety system or a substitute for human review in high-impact decisions. Weak at dense counting; the adapter route is specialised for screens and geometry; probabilities cover only the supplied options (no abstain answer); calibration reflects our data mix and may transfer imperfectly; evaluated on English questions with thinking mode off. Native-resolution cost and latency grow with image size.

Licence

Code, adapter and calibration: Apache-2.0. The base model Qwen/Qwen3.5-9B is Apache-2.0 and is not redistributed here. See LICENSE, BASE-MODEL-LICENSE and NOTICE; third-party data licences as listed above.