Hopper
A rank-16 LoRA on Qwen3.5-4B built for the JevBench setting. A document, a policy and a question go in; a calibrated distribution over the option letters comes out of one forward pass. Research and demo use only.
On the 231 public JevBench items, measured by the authors rather than the JevBench maintainer, it scores 100% easy, 94.4% standard and 68.5% hard (the frozen base with the same prompt: 63.1% hard), with a hard-tier calibration error of 0.102 under the 1.0.0 map. The authors disclose that half the public items served as a development gate; on the reserved half it is level with the frozen base. One pass weighs up to 26 options; serving code 1.1.1 handles longer menus in two stages (BANKING77 67.3% at 321 ms p50). The weights are for research and demo use only, because the training data includes RACE passages under non-commercial terms, so this registry links to them and does not mirror them. The serving code is Apache-2.0.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3.5-4b |
| Maker | HopitAI |
| Released | 2026-09-21 |
| License | Research and demo use only (training data includes RACE, non-commercial); serving code Apache-2.0 |
| Reported accuracy | 68.5% |
Get the weights
pip install systemonemodels
systemone pull hopit-ai/hopper
The files are served from the maker's Hugging Face repository, HopitAI/hopper, and verified against the checksums recorded here.
Read more
This page was opened by System One for HopitAI, who can claim the organisation and take it over at any time.