反向工程的Jev类模型开源实现发布

开源开发者发布了一个独立复现的Jev类模型实现1。该模型是对TypeSafe商业模型Jev的启动复现，设计方案未经官方公开1。不同于传统逐词生成的方式，该模型可以从给定的文本选项列表中进行单遍选择1。

在性能表现上，该模型在Wikispeedia数据集上通过冻结Qwen2.5-0.5B编码器实现了26%的准确率，相比随机基线的8%有显著提升1。独立训练的小型字节编码器在40,000次点击后达到29%准确率1。单遍评分方式的效率优势明显，相比被迫生成400个标记的小型解码器快约100倍1。在合成菜单数据上，该模型达到约98%的准确率1。

项目还包含了具体的应用示例。Doom游戏示例在10个记录的剧集中平均获得0.60击杀和-97.50奖励1。国际象棋示例的表现存在较大差异，对随机对手的50场比赛中为4胜46平0负，但对阵Stockfish 0级则为0胜2平48负1。该项目提供了详细的训练、评估和使用方法文档1。

An independent reverse-engineering effort has produced an open-source implementation of a Jev-like model, a text selection system designed to choose from predefined options in a single pass rather than generating responses word by word 1. The project represents a community attempt to replicate TypeSafe's commercial Jev model, whose design details remain undisclosed 1.

The model demonstrates competitive performance across multiple benchmarks 1. When freezing a Qwen2.5-0.5B encoder on the Wikispeedia dataset, it achieves 26% accuracy compared to 8% for random baseline selection 1. A smaller byte-level encoder trained from scratch over 40,000 clicks reaches 29% accuracy 1. On synthetic menu data, the model attains approximately 98% accuracy 1. The single-pass scoring approach proves roughly 100 times faster than forcing a small decoder to generate 400 tokens 1.

Practical applications included in the release demonstrate the model's versatility 1. A Doom game example averages 0.60 kills and −97.50 reward across 10 recorded episodes 1. A chess example achieves 4 wins and 46 draws against random players over 50 games, while showing 0 wins and 2 draws against Stockfish level 0 across 48 games 1. The project includes comprehensive documentation for training, evaluation, and deployment 1.