joebumodels/JevFinetune

Multimodal JEV

Multimodal JEV is a visual judgment model based on Qwen3.5-4B. It evaluates questions about images and returns structured judgments and probabilities rather than generating free-form answers.

Supported tasks

• choice: Select an answer from candidate options, with probabilities and an abstention probability.
• noul: Estimate the probability that a statement is true.
• score: Assess an ordered scale and return probabilities for its levels.

Model files

The checkpoint includes model weights, decision and ordinal heads, tokenizer, image processor, configuration and calibration metadata. Weights are stored in six indexed safetensors shards, totaling approximately 9.7 GiB.

Usage

Inference requires the custom UnifiedJev loader, the mso runtime modules and their Python dependencies. These must be available on your Python import path; downloading the checkpoint alone is not sufficient. Standard Qwen loading does not implement the JEV judgment interface.

from unified_loader import UnifiedJev

engine = UnifiedJev('/path/to/checkpoint')
answers = engine.system_one(
    {'images': ['/path/to/image.jpg']},
    {'q': {
        'type': 'choice',
        'instructions': 'Are tooth-colored attachments visible on any teeth in this image?',
        'options': [{'text': 'Yes'}, {'text': 'No'}],
    }},
)
print(answers)

The default image pixel budget is 262144. The loader uses BF16 for the backbone on CUDA and FP32 for the judgment heads. Set MSO_FLA=0 to use the reference attention implementation when the optional accelerated kernels are unavailable.

Comparison with prompt-based Qwen output

Comparison	Prompt-based Qwen output	Current JEV-style approach
Output method	Generates text or JSON token by token	Task heads directly compute probabilities
Probability source	Generated numbers, or separately computed answer-token probabilities	Decision heads compute probabilities from features
New tasks	New prompts can be tried immediately	New questions can be tried; specialized capabilities usually require labeled fine-tuning
Format stability	May require constrained generation and parsing	Fixed structured output
Explanations	Can generate explanations	Judgment heads do not generate explanations
Learning method	Prompts, examples, or generative fine-tuning	LoRA and task heads trained with labels and task losses
Six questions on one image: measured total latency	4,435 ms: generates all six answers as JSON	188 ms: jointly evaluates all six questions

Limitations

Returned probabilities are model estimates and may be incorrect or poorly calibrated. Performance depends on image quality, question wording and candidate options. Dental examples do not establish clinical reliability; this model is not a substitute for professional diagnosis.

Licenses

Refer to LICENSE-Qwen and LICENSE-OmniJev in the checkpoint directory for the applicable upstream licenses and notices.