philipp-zettl/vrom-hf-docs

vROM（Vector Read-Only Memory）是一个预计算的、序列化的HNSW索引包，可以直接加载到VecDB-WASM中，用于在浏览器中进行即时向量搜索，无需在客户端进行嵌入计算。该数据集包含了Hugging Face Transformers（v5.6版本）和Hugging Face Hub的预嵌入文档，涵盖了安装、快速开始、管道API、训练、微调、任务、量化、API参考等内容。数据集提供了1,356个向量，每个向量有384个维度，总标记数约为233K，索引大小为12.6 MB。嵌入模型使用的是Xenova/all-MiniLM-L6-v2（q8），距离度量采用余弦相似度。数据集还包含了详细的元数据结构和分块策略，以及如何在浏览器中使用VecDB-WASM进行快速开始的示例代码。

A **vROM (Vector Read-Only Memory)** is a pre-computed, serialized HNSW index package that can be loaded directly into [VecDB-WASM](https://huggingface.co/spaces/philipp-zettl/vecdb-wasm) for instant vector search in the browser — no embedding computation required on the client side. This vROM contains pre-embedded documentation from Hugging Face Transformers (v5.6) and Hugging Face Hub, covering installation, quick start, pipeline API, training, fine-tuning, tasks, quantization, API reference, and more. The dataset provides 1,356 vectors with 384 dimensions each, totaling approximately 233K tokens, and an index size of 12.6 MB. The embedding model used is `Xenova/all-MiniLM-L6-v2` (q8), with cosine distance metric. The dataset also includes detailed metadata schema and chunking strategy, along with example code for quick start in the browser using VecDB-WASM.