GestaltLabs/Jeff-1

Jeff 1

An open-source model for typed decisions, with local inference and a guide to training your own.

Jeff takes text or structured data and a set of questions, then returns labels and probabilities instead of generated prose. Label names and descriptions are supplied at call time.

• Model weights
• Source code
• Train your own Jeff
• Coding-agent guide
• License: Apache 2.0 for the code and adapter

What it does

Jeff supports three question types:

Type	Input	Output
choice	Named labels with descriptions	Selected label, probabilities, and confidence
noul	A yes/no question	Probability of yes
score	Ordered levels with descriptions	Level probabilities, expected level index, and confidence

The model is a LoRA adapter for Qwen3-4B-Instruct-2507. It scores candidate labels using the language model's token probabilities. It uses first-token scoring when candidates have distinct first tokens and whole-sequence scoring otherwise. Confidence is the largest label probability.

Why Jeff

Typed outputs let applications use model judgments without parsing generated text. Open weights let you inspect, adapt, and run the model locally; after the initial download, inference does not require a hosted model API.

Jeff supports Jev-style Choice, Noul, and Score requests. It is an independent project, not affiliated with TypeSafe AI. API compatibility does not imply identical judgments or performance.

Results

The released adapter was evaluated against live Jev 1.13.0 on the same 9,730 human-labelled fact-checking examples from FEVER, VitaminC, SciFact, and Climate-FEVER.

Model	Accuracy	Macro-F1	Brier ↓	ECE ↓
Jeff 1	0.8183 (7,962/9,730)	0.7789	0.2839	0.0807
Jev 1.13.0	0.8283 (8,059/9,730)	0.7994	0.2750	0.0932

Jeff has lower accuracy and Brier performance than Jev on this evaluation, but lower expected calibration error (ECE). Both ECE values use maximum class probability and ten equal-width bins. Jev's separate API confidence statistic is not used in this comparison. Lower ECE does not guarantee that an individual prediction is correct.

Jeff identifies supported claims more reliably than claims with insufficient evidence. The reliability plot compares predicted confidence with observed accuracy:

These results describe GestaltLabs/Jeff-1 (lora_4b_multi), not the older Choice-only adapter. This evaluation set has also been used for error analysis; it is not an untouched benchmark for future versions. Evaluation details and the recomputed metrics document the comparison.

Quickstart

Clone the full repository and install its dependencies with uv:

git clone https://github.com/Gestalt-Lab/jeff
cd jeff
uv sync

Run Python examples with uv run python. The first model load downloads the base model and adapter. Inference needs enough memory for the 4B base model; CUDA and Apple Silicon MPS are supported.

from jev_clf.client import SystemOneClient, Choice, Noul, Score

client = SystemOneClient(adapter="GestaltLabs/Jeff-1")
result = client.system_one(
    {
        "claim": "The museum opened in 1998.",
        "evidence": ["The museum first opened to visitors in 1998."],
    },
    {
        "verdict": Choice(
            instructions="Judge the claim using only the supplied evidence.",
            criteria={
                "supported": "The evidence establishes the claim.",
                "refuted": "The evidence contradicts the claim.",
                "not_enough_info": "The evidence is insufficient to decide.",
            },
        ),
        "has_date": Noul(instructions="Does the evidence include a year?"),
        "strength": Score(
            instructions="How strongly does the evidence settle the claim?",
            criteria=["none", "weak", "moderate", "strong"],
        ),
    },
)
print(result.choices["verdict"].choice)
print(result.choices["verdict"].probabilities)
print(result.nouls["has_date"].noul)
print(result.scores["strength"].score)

Local HTTP server

uv run python -m scripts.jev_clf_server

Open http://127.0.0.1:8079 for the demo. The server exposes POST /v1/systemone, GET /health, and GET /v1/models.

Load with PEFT

For lower-level access:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-4B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype="bfloat16")
model = PeftModel.from_pretrained(model, "GestaltLabs/Jeff-1").eval()

Load the tokenizer from the base model. PEFT loads the weights; the Jeff client adds prompt construction and label scoring in jev_clf/readout.py.

Train your own Jeff

The training walkthrough explains how to:

1. Represent your task as states, questions, and labelled answers.
2. Prepare separate training, validation, and test data.
3. Fine-tune a LoRA adapter on a Colab GPU.
4. Evaluate the adapter and load it with the same client.

It explains the learning objective and label-scoring mechanism as well as the commands. The small examples illustrate the format; a useful model needs a representative dataset and independent evaluation.

Limitations

• Unsupported positive verdicts. Jeff can mark a claim as supported when evidence supports only part of it or merely discusses the same topic. Claims combining several assertions are a known weakness.
• Insufficient evidence. Performance is weaker when the correct answer is not_enough_info, particularly with a single evidence passage.
• Evidence only. Jeff does not retrieve sources or independently establish whether supplied evidence is true. It can give confidently wrong answers.
• Per-question computation. Questions are scored separately. Labels that require whole-sequence scoring incur additional computation.
• Limited generalization evidence. The main evaluation is fact-checking. Small Choice, Noul, and Score checks verify supported interfaces, not broad reasoning equivalence to Jev or reliability on a new task.

Adapter details

This repository contains the adapter, not the base-model weights.

Property	Value
Base model	Qwen/Qwen3-4B-Instruct-2507
Adapter	LoRA, rank 16, alpha 32, dropout 0.05
Target modules	q_proj, k_proj, v_proj, o_proj
License	Apache 2.0

The release training configuration and available reproducibility information are documented in the model card. The base-model revision was not pinned during the release training run.

File	Purpose
adapter_model.safetensors	Adapter weights
adapter_config.json	PEFT adapter configuration
train_metrics.json	Recorded training metadata
SHA256SUMS.txt	Checksums
LICENSE / NOTICE	License and attribution

Adapter SHA256: 13cc3805495f7e901ca3121c7a3647fc6abcfe1fdc098ddc9ab1acd74f436a6a

Citation

@software{jeff1_2026,
  title  = {Jeff 1: a local open-weight typed decision model},
  author = {Gestalt-Lab},
  year   = {2026},
  url    = {https://github.com/Gestalt-Lab/jeff},
  note   = {Adapter: https://huggingface.co/GestaltLabs/Jeff-1}
}