Jev
Jev reads a JSON request with state and questions and returns option scores for choice, score, and noul. This export contains the merged Qwen2-VL backbone, Jev scorer, tokenizer, and inference code. It also includes the frozen vision encoder and image processor. Check the original Qwen license before redistributing.
from huggingface_hub import snapshot_download
import sys
path = snapshot_download(repo_id="bqbbao6/bjev")
sys.path.insert(0, path)
from jev.hub import Jev
model = Jev.from_pretrained(path, device="cuda")
message = '{"state":"I was charged twice.","questions":{"intent":{"type":"choice","instructions":"What does the customer want?","criteria":["refund","information"]}}}'
print(model.predict_json(message))
For image requests, add images with local paths or base64 data URIs, then pass the JSON string to predict_json. The pretrained vision encoder has not been fine-tuned on image examples.