jev-qwen35-08b

jev08 — Typed Decision Model (LoRA adapter)

Unofficial community fine-tune. Not affiliated with Alibaba/Qwen or the Open-Jev authors. Base weights are not redistributed — load Qwen/Qwen3.5-0.8B-Base (revision dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68) yourself.

A Jev-style System One decision model: prefill-only, never generates text. It scores each candidate answer with a scalar head over the last prompt token and returns temperature-calibrated typed probabilities for three question kinds: choice (single select), score (ordinal levels), noul (binary yes/no).

Architecture

• Frozen Qwen/Qwen3.5-0.8B-Base backbone + LoRA r=8 (α=2r, dropout 0) on attention/DeltaNet projection modules
• Scalar decision head Linear(hidden→1), initialized as W_lm["Yes"] − W_lm["No"] (the pretrained yes-direction), kept in fp32
• Candidates rendered as independent Yes/No judgment prompts (template byte-identical to Open-Jev); one batched forward per question; readout at the last non-pad token; max_length=768 (no silent truncation)
• Single-scalar temperature T=1.9190 fitted on 512 calibration rows only (NLL-minimizing), frozen for all evaluation

Training

Trained on ZefanCai/Open-Jev release-v2-redistributable (CC0): 19779 optimizer steps (AdamW, LoRA lr 5e-5 / head lr 1e-4, soft-label CE + 0.1·Brier, accumulation 4, seed 42), wall clock ≈ 4.6 h.

Metrics (held-out, never used for training or calibration)

See limitations in the training run summary: single seed, not an official Open-Jev reproduction; latency includes tokenization; synthetic OOD measures template/entity transfer, not world knowledge.

Usage

This repo ships a minimal jev08/ inference package (clone it, then):

from jev08.predict import predict

# request.json = {"state": {...}, "questions": {"q1": {"type": "choice", ...}}}
result_json = predict(checkpoint="path/to/adapter-dir", request_path="request.json")

Or drive the pieces manually:

from jev08.model import DecisionModel
from jev08.api import compile_request, format_response
from jev08.metrics import softmax
import json, torch

model = DecisionModel.load("path/to/adapter-dir")   # needs base model access
records = compile_request(state, questions)
temperature = json.load(open("path/to/adapter-dir/temperature.json"))["temperature"]
with torch.inference_mode():
    logits = model(records)
probs = [softmax(row.float().tolist(), temperature=temperature) for row in logits]
answers = format_response(records, probs)

The model never produces free text: output space is fixed at request compile time; the model only supplies probabilities, hard-validated (finite, in [0,1], sum to 1).

Licenses

See LICENSE-NOTICES: Apache 2.0 (base model, this adapter), MIT (ported Open-Jev code), CC0 (training data — no rows redistributed).