Qev-2B

Qev-2B is the compact member of the Qev decision-model family. It learns option probabilities and representation responses from Qev-9B, and supports Choice, Noul and Score with the same request format.

Model weights · Source and installation · 中文说明

from qev import Qev

model = Qev.from_pretrained("AustinFu/Qev-2B", device="cuda")

Install Qev from the source repository using Python 3.12 and a hardware-compatible PyTorch 2.8.0 build. The loader downloads this adaptation package and the compatible Qwen base separately. The teacher is not needed for inference. Use revision="v0.1.0" to select the first published release.

To fine-tune on your own data, use configs/qev-2b-finetune.json with python -m qev.train --init-checkpoint AustinFu/Qev-2B. To train with a teacher, follow the distillation guide or 中文指南.

Accuracy (%). Public JevBench accuracy is 172/231. These are the selected checkpoint's recorded results. The native base uses its language-model head and zero-shot prompts. The comparison therefore includes architecture, training and readout differences; it does not isolate the contribution of response distillation. Full results.

The package includes the LoRA adapter, decision head, interaction gate, tokenizer, model configuration, both training configurations, recorded evaluation results and license files. The three trained tensor files preserve the selected checkpoint exactly; optimizer state and base weights are excluded. Qev-train publishes 1,842 synthetic source examples with original hard labels and generation documentation. The full mixed training corpus, teacher probability caches and response-distillation data are not distributed.

Qev supports research and development of routing, rule judgments and rubric ratings over explicit options. Evaluate the model on your application's inputs; probability calibration and production reliability have not been established.

Qev adaptation weights use Apache-2.0. Source datasets and the Qwen base retain their own terms. License and attribution.