Jegue 0.1

Jegue is "donkey" in Portuguese. Version 0.1 is a 9B typed-decision model: given a question with options (choice or yes/no), it returns a calibrated probability for each option in a single forward pass, so it runs locally on one consumer GPU.

What it is

• Base model: Qwen/Qwen3.5-9B (bf16), plus a LoRA adapter (rank 32, alpha 64), a small answer-slot readout head and a small calibration head.
• Answers are read at an answer slot (the next token through the option rows of the LM head plus a trained residual); no text is generated and there is no reasoning step. Probabilities are temperature-calibrated per question.
• Prompts longer than 32,768 tokens are refused as unsupported and never truncated.
• Training method is not described in this release.

Training data note

The model was trained on public train/dev splits of upstream datasets plus synthetic examples. Every training example was screened against all public-suite rows (exact state match and 8-word-window overlap), and no public-suite row was used for training. Information about the public suite's distributions (unlabeled inputs and class proportions of some benchmarks) informed the design of the training; no public-suite labels or answers were used.

Decision Index 0.3, public suite (complete run, all requests answered)

Decision Index 54.13 (raw 65.39, breadth 51.78). 140,178 requests, none failed.

Area	Benchmarks	Raw	Chance-corrected
Knowledge & Reasoning	11	48.6	33.5
Language Understanding	9	74.0	62.5
Retrieval & Classification	6	74.2	65.5
Tools & Automation	5	74.5	72.3
Arts & Human Taste	6	52.2	29.8

Per-benchmark numbers: results/runs/jegue-0.1/benchmark-summary.json and index.json. Known weak spots: HLE, GPQA Diamond, ChessBench, POP909-CL, and long-prompt reasoning (CRUXEval, Habermas Machine).

Inference settings of the submitted run

• Engine: decision_engine_dense:DenseSlotEngine (in code/), options model=Qwen/Qwen3.5-9B, run=<weights>, ckpt=<weights>; LoRA merged into the bf16 weights, temperature head feat, prefix cache on, context limit 32,768 tokens.
• Hardware: one NVIDIA RTX 5090 (32 GB), torch 2.8.0+cu128, CUDA 12.8. Latency in that run: median 46.6 ms, mean 132.4 ms, p95 320.4 ms per request.

Try it

One CUDA GPU (about 20 GB), the Decision Index kit installed (the engine builds on its Engine class), and the base model Qwen/Qwen3.5-9B in the Hugging Face cache. From the repo root, put code/ on the path and ask a multiple-choice question and a yes/no question about a shared context:

import sys
sys.path.insert(0, "code")
from decision_engine_dense import DenseSlotEngine

engine = DenseSlotEngine(model="Qwen/Qwen3.5-9B", run="weights", ckpt="weights")

state = "Email from a customer: 'My parcel was due on Monday and it is Friday. Tracking has not changed since Tuesday. I want my money back.'"
questions = {
    "intent": {
        "type": "choice",
        "instructions": "What does the customer mainly want?",
        "criteria": {"refund": "A refund", "status": "An update on the delivery", "complaint": "To complain about the service", "other": "Something else"},
    },
    "urgent": {
        "type": "noul",   # yes/no: returns the probability of yes
        "instructions": "Should this email be escalated to a human agent today?",
    },
}

response, _ = engine(state, questions)
print(response["answers"]["intent"]["probabilities"])
print(response["answers"]["urgent"]["noul"])

Output of the full script (examples/ask.py, the same example with formatted printing; run it with PYTHONPATH=code python examples/ask.py):

intent: refund  refund=0.88  status=0.04  other=0.04  complaint=0.04
urgent: P(yes)=0.68

Run the benchmark

From a checkout of the Decision Index kit:

PYTHONPATH=<this repo>/code python -m decision_index pipeline \
  --engine decision_engine_dense:DenseSlotEngine \
  --option model=Qwen/Qwen3.5-9B --option run=<this repo>/weights --option ckpt=<this repo>/weights \
  --suite-dir <suite-0.3> --out runs/jegue-0.1

Layout

• weights/: lora/ adapter, head.pt (answer-slot residual), temperature.pt and temperature.json (calibration head), meta.json (step, base model).
• code/: inference engine and the modules it imports.
• results/runs/jegue-0.1/: the complete run (results.jsonl, scores.json, index.json, benchmark-summary.json, environment.json, status.json).
• examples/: ask.py.
• assets/: the logo.

License

Apache License 2.0 (see LICENSE), the same license as the base model Qwen/Qwen3.5-9B.