rev

What it is

rev is a decision model built on a causal LM backbone (Qwen/Qwen2.5-0.5B to 8B) with a LoRA adapter and bilinear pointer readout head. It reads a document once and evaluates multiple typed questions in parallel in a single prefill forward pass with zero autoregressive text generation. Weights are on Hugging Face Hub.

How it uses Jev

rev implements the TypeSafe System One API (POST /v1/systemone) and is compatible with the official typesafe-sdk. It takes a state plus typed questions (choice, noul, score) and returns calibrated probabilities. The model uses a pointer head to score options and softmax to produce typed answers and confidence scores.

Primitives:choicenoulscore

Technique worth stealing

Packing state and question branches with block-causal mask and branch position IDs for exact isolation and order invariance.

Try it

pip install -e . then python -m rev.serve --run jaswanthsanjay88/rev-0.5b --port 8000

View on GitHub