README.md

---
title: Vashtra
colorFrom: indigo
colorTo: purple
sdk: static
app_file: index.html
pinned: true
license: apache-2.0
short_description: A 0.6B ML-specialist model running fully in your browser
models:
  - vigneshwar234/Vashtra-0.6B
datasets:
  - vigneshwar234/vashtra-ml-corpus
tags:
  - webgpu
  - transformers.js
  - on-device
  - machine-learning
  - small-language-model
---

# Vashtra

A 0.6B model tuned for machine learning questions, running entirely in your browser
on WebGPU. There is no server and no API key. The weights download once and cache,
and every token after that is generated on your own GPU, so nothing you type leaves
the page.

Ask it about batch norm, LoRA against full fine-tuning, focal loss, exploding
gradients. Those are the kinds of questions it was trained on.

## How it works

The model is continued-pretrained on 117k arXiv `cs.LG` abstracts, then fine-tuned on
ML, stats and CS Q&A, with general chat data mixed in so it can still hold a normal
conversation. It is exported to ONNX, quantised, and loaded here with
[transformers.js](https://github.com/huggingface/transformers.js).

- Model: [vigneshwar234/Vashtra-0.6B](https://huggingface.co/vigneshwar234/Vashtra-0.6B)
- Corpus: [vigneshwar234/vashtra-ml-corpus](https://huggingface.co/datasets/vigneshwar234/vashtra-ml-corpus)
- Code: [github.com/vignesh2027/Vashtra](https://github.com/vignesh2027/Vashtra)

## What you need

A browser with WebGPU: Chrome or Edge 113 and up, or Safari 18 and up. The first load
pulls a few hundred MB. After that it starts from cache.

## Limits

It is a 0.6B model. It explains concepts reasonably and gets specifics wrong, so
treat exact numbers, named papers and anything recent with suspicion. Check answers
before relying on them.

Built by [vigneshwar234](https://huggingface.co/vigneshwar234).