README.md

metadata

license: mit
pretty_name: MoJev
base_model: Qwen/Qwen3.5-0.8B
datasets:
  - MoLeMo-Lab/mojev-mix
library_name: transformers
pipeline_tag: text-classification
tags:
  - calibration
  - structured-output
  - multiple-choice
  - preference-learning
  - multimodal
  - mojev
model-index:
  - name: MoJev
    results:
      - task:
          type: text-classification
          name: Typed decision scoring
        dataset:
          type: MoLeMo-Lab/mojev-mix
          name: MoJev-Mix test
          split: test
        metrics:
          - type: accuracy
            value: 0.9323
            name: Accuracy

MoJev

Typed, calibrated decisions in one forward pass.

Contact: contact@molemo.org

This is the trained checkpoint for the mojev runtime. It scores request-time candidate values from unstructured state and returns schema-bound probability distributions.

MoJev family resource	location
Code	MoLeMo-Lab/mojev
Model	MoLeMo-Lab/mojev
Dataset	MoLeMo-Lab/mojev-mix
Results	MoJev results
Preprint	MoJev (PDF)
Project page	MoJev

Interactive demo

Try MoJev on Hugging Face Spaces: text, one or multiple images, a question, and custom candidates are scored on server-side ZeroGPU. No model weights are downloaded to the browser.

The optional browser/ text export uses asymmetric INT4 linear weights, INT8 token embeddings, and an FP32 decision head. Export scripts and numerical comparisons are described in the browser guide.

Model contract

input	released configuration
state	text and local image references; 16,384-token training truncation
question	instruction text
candidates	request-time strings
output	logits decoded as Choice, Noul, or Score distributions

Candidate names are supplied by the caller and encoded directly from their text.

MoJev's Qwen3.5 backbone supports 262,144 tokens natively and up to 1,010,000 tokens with YaRN scaling. The checkpoint records the 16,384-token training window. The MoJev runtime accepts a larger inference state window through --context-tokens; the packed sequence also includes question and candidate tokens.

Run with the MoJev server

git clone https://github.com/MoLeMo-Lab/mojev
cd mojev
pip install -e '.[transformers]'

mojev serve MoLeMo-Lab/mojev --port 8000

For a longer state within the native context:

mojev serve MoLeMo-Lab/mojev --port 8000 --context-tokens 65536

from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8000") as client:
    result = client.system_one(
        state={"document": "I was charged twice. Please fix this ASAP."},
        questions={
            "category": Choice(
                instructions="What is this ticket about?",
                criteria={"billing": None, "technical": None, "other": None},
            )
        },
    )

print(result.choices["category"].choice)
print(result.choices["category"].probabilities)

Load with Transformers

from transformers import AutoModel, AutoProcessor

model = AutoModel.from_pretrained(
    "MoLeMo-Lab/mojev",
    trust_remote_code=True,
).to("cuda").eval()
processor = AutoProcessor.from_pretrained(
    "MoLeMo-Lab/mojev",
    trust_remote_code=True,
)

The model directory contains the scorer implementation through auto_map. Packing, candidate sorting, and schema decoding are provided by the mojev package.

Multimodal input

pip install -e '.[transformers]'
mojev serve MoLeMo-Lab/mojev --port 8000

Use the image marker followed by an absolute path in the request state:

from pathlib import Path

image = Path("examples/cat.jpg").resolve()
state = f"Identify the subject. <|vision_start|><|image_pad|><|vision_end|>{image}"

The processor expands the image into visual patch tokens in the state branch. Every question and candidate in the request can attend to those tokens.

candidate set	grey image P(cat)	cat image P(cat)
cat, dog	0.471	0.786
cat, dog, car, other	0.264	0.528

Evaluation

On 12,000 evaluation decisions, the released checkpoint reaches 93.23% accuracy with 0.79% expected calibration error.

Architecture

• Base: Qwen3.5-0.8B; all 854,036,544 parameters trained.
• Packing: state, questions, and candidates in one sequence.
• Attention: tree mask with isolated question/candidate branches.
• Readout: rank-512 context and candidate projections with scaled dot product.
• Objective: Plackett–Luce ranking plus Brier calibration loss.
• Precision: bf16 encoder and fp32 readout.

Training

item	value
data	205,084 rows from 18 Open-Jev generators
epochs	1
parallelism	8-way data parallel
learning rate	1e-5
state truncation	16,384 tokens
Brier weight	1.0
wall time	47 minutes

Applications

• routing and triage;
• policy and evidence classification;
• tool and workflow selection;
• calibrated execution, deferral, and escalation thresholds;
• multiple typed decisions over shared state.

Code is MIT licensed. The Qwen base model license applies to the checkpoint.