0.6B 模型在 MTEB(Law) 排名第二：高效法律嵌入的配方

AI 摘要

Hugging Face 博客发布了一篇技术文章，介绍了一个名为 dinghy-law 的 0.6B 参数法律嵌入模型，在 MTEB(Law) 基准上排名第二，超越了更大的模型。文章详细描述了在时间和计算资源有限的情况下，通过 LoRA 微调、基于梯度对齐的硬负样本挖掘、数据读取和权重空间合并等技术，实现了高性能。该模型基于 Qwen3-Embedding-0.6B，通过一系列实验验证了方法的有效性。

正文节选

| Model | Params | MTEB(Law) Mean(Task) | |---|---|---| | Euler-Legal-Embedding | 7–8B | ~70.4 | | ours (dinghy-law) | 0.6B | ~65.9 | | voyage-law-2 | — | ~65.4 | | F2LLM-8B | 8B | ~63.5 | | F2LLM-4B | 4B | ~61.5 | We had a target to hit, and two things in short supply getting there: time and compute. This is the recipe we built to hit it under those constraints. With more of either you'd make different calls — this is just what worked for us. Below is each step and why it's in the recipe.

发布时间：—

抓取时间：2026-08-09 18:22

来源机构：Hugging Face