philipp-zettl/vrom-ml-training

🧩 vROM: ML Training Stack (TRL + PEFT + Datasets)

Vector Read-Only Memory — Pre-computed HNSW index for instant in-browser RAG

What is this?

A plug-and-play RAG cartridge containing pre-embedded documentation for the ML training stack:

• TRL — SFT, DPO, GRPO, PPO, Reward, KTO, ORPO, CPO trainers
• PEFT — LoRA, adapters, parameter-efficient fine-tuning
• Datasets — Loading, processing, streaming, creating, uploading

Load directly into VecDB-WASM for instant vector search — zero compute embedding required on the client.

Metric	Value
Vectors	629
Dimensions	384
Total Tokens	~100K
Index Size	5.8 MB
Embedding Model	Xenova/all-MiniLM-L6-v2 (q8)
Distance Metric	Cosine

Quick Start

import init, { VectorDB } from 'vecdb-wasm';
import { pipeline } from '@huggingface/transformers';

await init();

// Load the vROM (5.8 MB)
const resp = await fetch(
  'https://huggingface.co/datasets/philipp-zettl/vrom-ml-training/resolve/main/index.json'
);
const db = VectorDB.load(await resp.text());

// Embed & search
const extractor = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2', { dtype: 'q8' });
const emb = await extractor('how to train with DPO', { pooling: 'mean', normalize: true });
const results = JSON.parse(db.search(new Float32Array(emb.data), 5));

Files

File	Size	Description
index.json	5.8 MB	HNSW index (VectorDB.load())
chunks.json	626 KB	Chunk metadata array
manifest.json	1.2 KB	Package spec
tools/vrom_builder.py	25 KB	Builder tool for custom vROMs

Part of the vROM Ecosystem

See also: vrom-hf-docs (Transformers + Hub docs)

Built with VecDB-WASM.