duguying

duguying / 100M

Public

training your 0.1B llm from zero

tinyllm.duguying.net llm training
29
4
100% credibility
Found May 08, 2026 at 29 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A user-friendly web studio for training a tiny Chinese language model from scratch on Apple Silicon Macs, covering data gathering, tokenization, pretraining, fine-tuning, and Ollama export.

How It Works

1
🔍 Discover Tiny LLM Studio

You hear about a simple web tool to train your own tiny Chinese AI helper right on your MacBook.

2
🚀 Open the web studio

Launch the friendly app in your browser to start building your personal language model.

3
📚 Gather Chinese texts

Easily collect stories, articles, and conversations from safe online sources into your library.

4
🔤 Create your word breaker

Train a custom tool that understands Chinese words perfectly for your texts – quick and magical!

5
🏗️ Prepare training sets

Organize your texts into ready-to-learn packs, splitting some for checking progress.

6
🧠 Teach it to chat

First teach basic language patterns, then conversations, making it respond like a helpful friend.

💬 Chat with your AI

Enjoy talking to your homemade Chinese assistant, and share it easily with friends via a simple app.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 29 to 29 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is 100M?

This Python repo lets you train a 0.1B parameter Llama-style LLM from scratch on Apple Silicon Macs using MPS acceleration. It covers the full pipeline: collect github training data from HF datasets or Wikipedia via a web UI, train a 32k BPE tokenizer, pretrain on up to 5B tokens (quick 30min test to 14-day scale), add SFT/DPO alignment, then chat or export GGUF for Ollama pulls. Solves laptop-bound github training ai without cloud costs, like a 100m weltrekord sprint for custom Chinese models.

Why is it gaining traction?

The killer hook is Tiny LLM Studio, a FastAPI web app for browser-based corpus jobs, dataset packing, training queues, and progress streams – far easier than scripting github training data flows. bf16 MPS configs hit 4K tok/s on M4 Pro for overnight pretrains, beating CPU alternatives. Ollama export means instant local inference, dodging github 100mb limit woes.

Who should use this?

Mac devs building Chinese chatbots from github training repo scraps. Hobbyists fine-tuning on private data sans GPUs, like alpaca-zh SFT in an hour. Proto-researchers testing DPO alignments locally before scaling.

Verdict

Grab it for fast 100M LLM experiments on M-series – web UI shines for iteration. But 29 stars and 1.0% credibility score mean it's raw: lean docs, no broad tests. Fork and tweak if github training online feels bloated.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.