jev-qwen3-0.6b

A choice-scoring model trained on avbiswas/bev-decision-150K. It uses the first 20 layers of Qwen/Qwen3-0.6B (LoRA r=8 on the last 8) and an attention head that scores the answer choices.

It answers three kinds of question about a text: pick one of several choices, yes/no (noul, returns P(yes)), and score on an ordered scale. The choices share position IDs and can't attend to each other, so the prediction doesn't depend on the order of the options.

This repo has the LoRA adapter (lora/), the head (head.safetensors), the settings (bev_config.json), and the model code (bev_model.py). The Qwen base weights are downloaded from their own repo.

Usage

# pip install torch transformers peft safetensors huggingface_hub
import sys
from huggingface_hub import snapshot_download

path = snapshot_download("NaheedRayan/jev-qwen3-0.6b")
sys.path.insert(0, path)
from bev_model import load_bev, answer

network, meta = load_bev(path)
state = "The package arrived two weeks late and the box was crushed, but support refunded me quickly."

answer(network, meta, state, {"type": "noul", "instructions": "Is the customer satisfied with the delivery?"})
answer(network, meta, state, {"type": "choice", "instructions": "Which team should handle this?",
                               "criteria": {"shipping": "delivery problems", "billing": "payments and refunds", "tech": ""}})
answer(network, meta, state, {"type": "score", "instructions": "How angry is the customer?",
                               "criteria": ["calm", "annoyed", "furious"]})

Results (step 3400)

metric	value
loss	0.5779
accuracy	0.7580
accuracy_noul	0.8475
accuracy_score	0.6345
accuracy_choice	0.7397
auc_noul	0.9103

Training

• State truncated to 512 tokens; choices to 64 tokens.
• Head: 2 attention blocks, width 512.
• Trained with the all-in-one Kaggle notebook from the bev-train repo.