First Instinct

A language model that chooses from the answers you define.

Give it a situation, a question, and the allowed answers. First Instinct scores those answers and returns a probability distribution. You supply the question and answer definitions on each request.

Download the model · Run the demo · Results · Documentation

The interface

{
  "state": "Expedite parcels that are at least two days late. This parcel is three days late.",
  "questions": {
    "next_action": {
      "type": "choice",
      "instructions": "Which action follows the policy?",
      "criteria": {
        "expedite": "Expedite the parcel",
        "wait": "Wait"
      }
    }
  }
}

The response contains the selected answer and probabilities for each option. The same model supports choices, yes/no questions, and ordered scores; see examples of all three.

Each question takes one model pass, with no generated explanation. Your application executes any chosen action. Selection probabilities describe the offered answers; they are not guarantees that an action will succeed.

Run the demo

The browser demo is Snake, played by the model. Each turn supplies the board and three directions, executes the selected move, and displays its probabilities. You can pause, advance one move, or inspect the exact request. The game supplies immediate collision observations; it does not override the model's choices.

Download and verify the archive from the model release, then extract it and run its verify.py. Use the current repository checkout for Snake; the release archive contains an earlier interface.

With your Python environment active, run from this checkout:

python -m pip install -r requirements-scale.txt
python -m general_lab.serve \
  --run /path/to/extracted-release/runs/supervised --device auto

Open localhost:8766. Foundation weights download on first use. Check the hardware requirements and setup guide before loading the 9B model.

To ask your own questions, edit the example file and run:

python -m general_lab.interface \
  --run /path/to/extracted-release/runs/supervised \
  --input examples/general-decisions.json --device auto

Model and results

The released checkpoint adapts Qwen3.5-9B on 350,857 examples using low-rank adapters throughout the language network. Training updates 43.3 million adapter parameters while keeping the original foundation matrices frozen.

Evaluation	Questions	Original Qwen	First Instinct
Full test mixture	17,277	63.3%	78.1%
Public sources held out from adaptation	13,077	66.2%	73.5%
Reserved executable task combinations	4,000	50.4%	71.7%

Values are question-weighted accuracy using the same constrained-answer interface for both models, without generated reasoning. The public-source rows are a subset of the full test; reserved combinations belong to a separate challenge set. Some tasks regress. The full report includes task breakdowns, probability scores, uncertainty, and known failures.

The demo uses this supervised checkpoint. Reinforcement-learning experiments have not yet qualified a replacement. Ongoing work tests whether learning from executed actions and verified outcomes improves decisions and consequence forecasts. See the experiment guide for protocols and results.

This is an independent, Jev-inspired research project. General calibration and equivalence to Jev have not been established.

---

Project guide · Environments and data · Reproduction and tests

Code: MIT. Data and upstream dependencies: third-party notices.