Official JevAny LoRA checkpoint using the pointer readout on
Qwen/Qwen3.5-4B. It requires the JevAny code
at the release revision linked from the project repository.
All values below use the same frozen Transfer and public JevBench protocols. They are accuracy, not the sealed JevBench leaderboard composite.
| Transfer (1,046) | JevBench Easy (48) | Original (72) | Hard (111) | JevBench total (231) |
|---|---|---|---|---|
| 78.68% | 100.00% | 95.83% | 61.26% | 80.09% |
Training used 1,772,725 text records / 2,180,242 labelled decisions. It spans preference, agent/tool decisions, reasoning, classification, and safety.
This checkpoint uses the pointer readout and supports more than 255 choices (subject to context limits). Direct-token models train with full-vocabulary cross-entropy and are therefore slower to train; pointer models score option representations with a small learned head. Inference speed is expected to be similar because both use a single backbone prefill and do not autoregressively generate an answer.
pip install -e '.[serve,multimodal]'
jevany serve --checkpoint SimpleJev/JevAny-Qwen3.5-4B-LoRA --device cuda --dtype bf16
The repository contains LoRA adapter weights and JevAny readout metadata, not the base weights. The base model's license and access terms also apply.
Official JevAny LoRA checkpoint using the pointer readout on
Qwen/Qwen3.5-4B. It requires the JevAny code
at the release revision linked from the project repository.
All values below use the same frozen Transfer and public JevBench protocols. They are accuracy, not the sealed JevBench leaderboard composite.
| Transfer (1,046) | JevBench Easy (48) | Original (72) | Hard (111) | JevBench total (231) |
|---|---|---|---|---|
| 78.68% | 100.00% | 95.83% | 61.26% | 80.09% |
Training used 1,772,725 text records / 2,180,242 labelled decisions. It spans preference, agent/tool decisions, reasoning, classification, and safety.
This checkpoint uses the pointer readout and supports more than 255 choices (subject to context limits). Direct-token models train with full-vocabulary cross-entropy and are therefore slower to train; pointer models score option representations with a small learned head. Inference speed is expected to be similar because both use a single backbone prefill and do not autoregressively generate an answer.
pip install -e '.[serve,multimodal]'
jevany serve --checkpoint SimpleJev/JevAny-Qwen3.5-4B-LoRA --device cuda --dtype bf16
The repository contains LoRA adapter weights and JevAny readout metadata, not the base weights. The base model's license and access terms also apply.