SelfJev-4B

SelfJev turns text and questions into structured decisions: yes/no probabilities, one choice, an ordered score, or every option that applies. It runs on your own GPU and exposes a Python SDK and HTTP API.

This repository contains the trained LoRA adapter (230 MB), its PEFT configuration, and its model manifest. The base model is Qwen/Qwen3.5-4B, pinned to revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a; the base weights download separately.

Source code · API reference · Research notebook

Use it

SelfJev requires its custom serving engine. It reuses the document computation across questions and scores answer choices directly, with no generated text. A generic Transformers text-generation or text-classification pipeline does not reproduce this API or the reported scores.

On a Linux machine with an NVIDIA GPU, Git and uv installed:

GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/Jwuthri/SelfJev.git
cd SelfJev
uv sync --frozen --no-dev --extra serve --extra gpu
uv run --no-sync python - <<'PYTHON'
from huggingface_hub import snapshot_download
snapshot_download(
    "JulienHeysam/selfjev-4b",
    local_dir="weights/selfjev_4b_hf",
    allow_patterns=["adapter_model.safetensors", "adapter_config.json", "model.json", "README.md"],
)
PYTHON

# Choose a secret for your own server.
export SELFJEV_API_KEYS="replace-with-your-long-random-secret"
uv run --no-sync selfjev serve --adapter weights/selfjev_4b_hf --host 127.0.0.1 --port 8000

In a separate terminal, use the same secret with the client:

from selfjev import SelfJev, Noul, Choice

client = SelfJev(
    base_url="http://127.0.0.1:8000",
    api_key="replace-with-your-long-random-secret",
)
result = client.system_one(
    state="I was charged twice. Please refund the duplicate payment.",
    questions={
        "refund": Noul("Does the customer want a refund?"),
        "team": Choice("Which team should handle this?", {
            "billing": "payments and refunds",
            "support": "technical issues",
        }),
    },
)
print(result.nouls["refund"].noul)
print(result.choices["team"].choice)

A 24 GB NVIDIA GPU is a practical starting recommendation, not a measured minimum. Plan for 16–32 GB host RAM and at least 50 GB free disk. Current-model GPU latency and a full-model CPU minimum have not been established. See the deployment guide for Docker and AWS.

Architecture and training

• Base: Qwen3.5-4B, with a rank-64 LoRA on attention and DeltaNet projections.
• Shared-prefix tree: read the document once, then evaluate questions and their answer choices in separate branches.
• Training: 79,943 non-test questions, one epoch, learning rate 2e-4, texts up to 16K tokens.
• Targets: 50% checked labels and 50% stored Jev probabilities as a teacher signal. Authored labels were checked by an AI judge; Jev predictions did not determine authored labels.
• All answer options are included in each question. The serving engine applies this transformation automatically.

See model.json for the original manifest, checkpoint and provenance. Independent project; not affiliated with TypeSafe or Qwen.

Recorded results

These are results from the current TreeServer engine:

Task	Questions	Accuracy	Evidence
Text decisions	1,991	95.7%	report
AI response review	946	93.1%	report
Broader text tasks (development benchmark)	3,471	83.8%	report

Accuracy is the share of questions matching the expected answer. For select-all questions, the entire set must match. The two focused test sets were authored and checked by AI; their labels are not human ground truth. They have informed research direction, and a fresh final holdout remains necessary. The broader text benchmark was repeatedly used during development. Scores are single-run observations, not guarantees on application data.

The same adapter scored 95.8% on text decisions under the original evaluation engine. model.json preserves those original scores; the table above reports current serving-engine measurements. Historic latency results for earlier Qwen3 prototypes do not establish current SelfJev-4B latency.

Artifact integrity

adapter_model.safetensors SHA-256:

dfbf2834d883987893ec305a6093a345fd79c77f903f60f2f914cbe3b6058d1b

The adapter uses the Qwen base model linked above. No separate adapter license is declared in this repository.