| { |
| "name": "selfjev-4b", |
| "release": "vision (images v1), 2026-09-30: the default since then; the text-only release is Jwuthrich/selfjev-4b", |
| "hub": "Jwuthrich/selfjev-4b-vision", |
| "base": "Qwen/Qwen3.5-4B", |
| "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a", |
| "recipe": "selfjev-4b (Jwuthrich/selfjev-4b) + selfjev finetune --lr 5e-5 --soft-weight 0.5 --max-length 16384, 1 epoch, 622 steps: 11,344 image questions from 6 licence-safe Hugging Face datasets (scripts/data/build_images_v1.py) + 11,344 text questions replayed from selfjev-4b's own training rows with Jev's probabilities as half-weight soft targets. The Qwen3.5 vision tower is frozen and unchanged; the LoRA (r64, attention + DeltaNet projections) is the only trained part.", |
| "prompt": "challenger-state-first-v1 (selfjev.engine.qwen35.Qwen35Scorer.entry)", |
| "options_in_question": true, |
| "inputs": "text, JSON, an image (base64 data URL) or a list of text and image parts", |
| "eval2": 96.1, |
| "dev_benchmark": 84.1, |
| "eval_llm": 92.5, |
| "eval2_by_type": { |
| "binary": 97.4, |
| "multiclass": 97.3, |
| "multilabel": 90.8 |
| }, |
| "eval_images_v1": { |
| "all": 90.4, |
| "trained_datasets": 94.2, |
| "held_out_datasets": 84.3 |
| }, |
| "serve": [ |
| "selfjev serve --adapter Jwuthrich/selfjev-4b-vision" |
| ], |
| "adapter_sha256": "5b619930d54ef041e4424d63b9a5078eb7b0a2d0b0535efec3a335ea33750179", |
| "source_run": "runs/images_v1/adapter", |
| "train_meta": "reports/train_meta/images_v1.json", |
| "reports": "reports/images_v1/", |
| "training_data_licences": { |
| "timm/oxford-iiit-pet": "CC BY-SA 4.0", |
| "zalando-datasets/fashion_mnist": "MIT", |
| "AI-Lab-Makerere/beans": "MIT", |
| "nateraw/rice-image-dataset": "CC0 1.0", |
| "timm/eurosat-rgb": "MIT", |
| "garythung/trashnet": "MIT" |
| } |
| } |
|
|
