Reverse-engineered Jev-like model

A compact starter model implements a Jev-like one-pass scorer that takes a text context and a variable-length list of text options and returns a probability for each option in a single forward pass. Each option is encoded into a query vector that attends over context tokens to produce an attended context vector; a shared dot product between option and context yields a score and a softmax over options produces choices. The project includes examples for text tasks and for visual controller scoring (Doom and chess demos), with supplied checkpoints: a joint checkpoint that averaged 0.60 kills and −97.50 reward across ten Doom episodes, and a chess-only checkpoint that scored 4 wins/46 draws/0 losses versus a random mover but 0/2/48 versus Stockfish level 0. Data rows are JSONL objects with context, a list of options and a zero-based label; default byte encoders truncate context to 192 bytes and options to 32 bytes.

Training and inference support a built-from-scratch byte encoder or a frozen Hugging Face encoder (example: Qwen/Qwen2.5-0.5B). Experiments show ~98% accuracy on synthetic menus; on target-disjoint Wikispeedia next-click data a frozen Qwen2.5 encoder plus the scorer reached 26% (controls ~8%), and a small model trained on 40k clicks reached 29%. One-pass scoring is roughly 100× faster than a small decoder writing 400 tokens at eight options. Limitations are explicit: this is a research starter (not a Jev reimplementation), byte encoding is limited for semantics, pretrained paths require downloads and memory, one-pass needs the full option list, and results depend on dataset and splits. Code is MIT-licensed.