jev-multimodal

An experimental Gemma 4 E4B instruction model LoRA adapter for typed decisions from text and up to four ordered images. It is inspired by Jev's noul (yes/no), choice, and score interaction pattern. This is an independent prototype; it is not the Jev model and has no affiliation with its creators.

This repository contains only the native MLX-VLM adapter files (adapters.safetensors and adapter_config.json). It does not contain the approximately 5 GB base checkpoint, a merged model, or a PEFT/Transformers adapter. The adapter needs the matching 4-bit base and the decision prompt/scoring code in the project source.

Base and adapter

Item	Value
Base	mlx-community/gemma-4-e4b-it-4bit
Base revision	475b9088d29754a3379866cf5aeb6b41acd313c2
Original model	google/gemma-4-E4B-it
Adapter format	MLX-VLM LoRA, rank 8, scale 2 (alpha 16), dropout 0
Tuned modules	Query and output projections in language layers 34–41
Weight SHA-256	87ef1029610d23688512883700c35aefa60f751850e8483225a64354701c3e45
Experiment	decision-v1, selected 2,000-step checkpoint

The base model's own model card and terms apply when obtaining and using its weights. This adapter was trained with Python 3.11, MLX-VLM 0.7.3, and MLX 0.32.2 on Apple Silicon/macOS.

Use

Download the pinned base and adapter into separate folders:

hf download mlx-community/gemma-4-e4b-it-4bit \
  --revision 475b9088d29754a3379866cf5aeb6b41acd313c2 \
  --local-dir models/gemma-4-e4b-it-4bit
hf download mohdyaser/jev-multimodal --local-dir models/jev-multimodal

From the project source, with its dependencies installed, a request JSON can be scored with:

.venv/bin/python src/jev_cli.py \
  --model models/gemma-4-e4b-it-4bit \
  --adapter models/jev-multimodal \
  --temperature 1.4914477134397612 \
  --input request.json

request.json contains a question, optional context, an ordered list of 0–4 image paths, an output_type (noul, choice, or score), and 2–16 { "id", "description" } options. For noul, option IDs are false and true; score may also specify increasing level_values. See the project README for the local UI and examples. The code builds the decision prompt, obtains the next-token logits for one-token option markers A–P in a single forward pass, then normalizes across the supplied options. The reported probabilities are conditional on those options; they are not confidence guarantees. The temperature above was fitted on the experiment's 100-example synthetic calibration split.

Training and evaluation

The frozen 4-bit base was adapted for one epoch on 2,000 locally generated synthetic decisions (0–4 images), with 250 optimizer updates. A 100-example development split selected this checkpoint. An independent 100-example split fitted temperature, and a 200-example synthetic test split measured the following:

Metric	Frozen base	This adapter
Decision accuracy	158/200 (79%)	158/200 (79%)
Calibrated negative log likelihood	0.551	0.494
Calibrated Brier score	0.289	0.272
Score mean absolute error (63 cases)	0.549	0.487
Paired image probes	6/9	6/9

The adapter improved probability quality on this small synthetic test but did not improve answer accuracy or paired image probe performance. Three- and four-image cases remained weak. All evaluated images were synthetic; no rights-checked natural-image test was included. These results do not establish performance on real photos, documents, dense OCR, unfamiliar prompts, or Jev-equivalent behavior. See the experiment report for the protocol and per-category results.