a local Jev-style decision head onto Qwen 2.5 1.5B

Jev from TypeSafe AI got me thinking about non-autoregressive decision engines, so I built a small proof-of-concept called AES to test the concept locally on open weights.

It uses Qwen/Qwen2.5-1.5B-Instruct as the base. Instead of generating tokens, it runs a single prefill pass, pulls the final hidden state through a custom FP32 SwiGLU probe, and calculates cosine similarity directly against candidate token embeddings. It implements the same NOUL, CHOICE, and SCORE decision primitives.

Full disclosure: I only trained it on a few datasets (~3,200 samples from BoolQ, ARC, and Yelp) for 1 epoch on free Kaggle T4 GPUs, so do not expect broad zero-shot generalization. It is strictly a small experiment.

It runs in ~28ms batched and works well on simple, unambiguous multiple-choice problems as long as you stick to single-token anchors (A, B, C, D or true/false) rather than long multi-word strings.

Hugging Face repo: https://huggingface.co/Brinij/aes