Jwuthrich/selfjev-4b-vision-merged

SelfJev-4B Vision — merged weights

Structured decisions from text and images. A 4B backbone. Your own GPU.

The SelfJev-4B Vision LoRA merged into the pinned Qwen3.5-4B checkpoint: the complete weights (about 9.3 GB), tokenizer and configuration, in the original multimodal layout. Only the language-model tensors the LoRA touches are replaced; the vision encoder is Qwen3.5-4B's own, unchanged. Provenance: merge_meta.json.

Adapter release (recommended) · Source code

Which one to use

• Images, or many questions per request: the adapter release with SelfJev's default engine (selfjev serve --adapter Jwuthrich/selfjev-4b-vision). The document or image is read once and every question branches from it.
• Text, one question at a time, fastest: this merged checkpoint on SelfJev's vLLM backend. The vLLM backend takes text only: SelfJev refuses image requests on it.

Quickstart (vLLM backend, text)

Use a Linux NVIDIA GPU machine with a vLLM environment that supports Qwen3.5.

GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/Jwuthri/SelfJev.git && cd SelfJev
python -m pip install -e '.[serve]'
huggingface-cli download Jwuthrich/selfjev-4b-vision-merged --local-dir selfjev-4b-vision-merged
selfjev serve --engine vllm --model-dir selfjev-4b-vision-merged

Results

Reference results of the source adapter through SelfJev's default engine (report files in the GitHub repository, reports/images_v1/). This merged artifact has not been benchmarked separately through vLLM.

evaluation	questions	SelfJev-4B Vision	Jev API
Text decisions (eval2)	1,991	96.1%	97.2%
AI response review (eval_llm)	946	92.5%	92.5%
Development benchmark	3,471	84.1%	82.7%
Image test (10 datasets, through the adapter release)	2,002	90.4%	no image input

The text-only release and its merged weights stay at Jwuthrich/selfjev-4b and Jwuthrich/selfjev-4b-merged. Training data, licences and limitations: see the adapter card.

Compared with open Jev-like models

Measured 2026-09-30. Eight open "Jev-like" decision models from Hugging Face and SelfJev were each served with their card's recommended setup on the same GPU (one NVIDIA L40S) and asked exactly the requests TypeSafe's Jev receives. A request a model cannot answer counts as wrong. Jev is shown for reference.

SelfJev's test suites. Text Decisions (1,991 questions) and AI Response Review (946), published as selfjev-decision-bench, and photos:

model	size	images	licence	Text Decisions	AI Response Review	held-out images	Text Decisions time (s)	AI Response Review time (s)
Jev 1.13 (TypeSafe API, reference)	?	no	paid API	97.2	92.5	—	—	—
openjev	27B	yes	CC-BY-NC-4.0	96.8	92.6	82.6	753	329
SelfJev-4B Vision	4B	yes	Apache-2.0 code	94.7	90.7	85.1	100	41
SelfJev-4B Vision, version 0.4.1	4B	yes	Apache-2.0 code	94.7	90.7	85.1	167	68
Plumb-4B	4.2B	no	Apache-2.0	93.4	84.1	—	471	164
jpt-4b	4.5B	yes	CC-BY-NC-4.0	92.5	85.0	83.8	90	40
imajev-4b	4B	yes	Apache-2.0	91.5	84.9	85.2	489	155
decider-4b	4.2B	no	Apache-2.0	90.8	82.8	—	145	52
Mica-v0.1-4B	4.2B	no	Apache-2.0	89.6 ²	84.1 ²	—	162	69
kev-4b	4B	no	Apache-2.0	88.4	74.7	—	109	34
Laya	0.4B	no	Apache-2.0	45.4	46.4	—	23	11

• Every model gets the same Jev-shaped questions; select-all questions become one yes/no per option, which costs SelfJev 1.4–1.8 points against its native evaluation (96.1 and 92.5). Held-out images: the 4 image datasets of the SelfJev image test that none of these models trained on (772 questions). Times: each suite end to end, 4 requests in flight; SelfJev 0.4.2 was measured 2026-10-05 on the same GPU model, and the row for version 0.4.1 is the same weights on the engine before it. ² Refuses inputs over 8,192 tokens (2 % and 1 % of the questions), counted wrong.
• No other model trained on these suites, but they come from the same authors and judges as SelfJev's training data, so they favour SelfJev.
• SelfJev is the most accurate open model at 4.5B parameters or less on both suites (paired tests, p ≤ 0.023) and ties imajev-4b on held-out images; the 27B openjev is higher on Text Decisions. Among the 4B models only jpt-4b is faster on Text Decisions, and kev-4b and jpt-4b on AI Response Review.

Protocol, per-model setups, paired tests and raw reports: reports/competitors.