FareedKhan-dev/train-llm-from-scratch

A straightforward method for training your LLM, from downloading data to generating text.

Language modelsPython#gemini#large-language-models#llm#openai#training

5.3 Breakout Magnitude out of 10. Star growth looks organic. Data as of September 20, 2026.

About the project

Educational project: a from-scratch PyTorch Transformer plus a full LLM training pipeline — from raw text through SFT, reward modeling, PPO/DPO/GRPO to chat, without transformers, trl or peft.

Useful for

Train a small LLM on a single GPU following the README steps
Study attention and Transformer internals from the source code
Walk through post-training stages: SFT, reward model, DPO, PPO, GRPO

README summarized by DeepSeek V4.1 Flash. Details may be inaccurate.

Why it’s trending

39 stars so far today, about 116 expected by the end of the day. The usual pace is 12 per day, so that's 9.3× as much.
Over the last two days the pace is 18× that of the previous week and a half.
The spike has held for 3 days in a row — not a one-off blip.
GitHub Trending Python today: #5, +196 stars.
About 40 forks a day — people are taking the code.

Stars per day

Bars are daily stars, the line is the usual pace. Red marks spike days.

Numbers

Total stars: 10,004
Stars in a day: 116
Forks: 1,389
Issues and pull requests: 40
Watchers: 80
Language: Python
License: MIT
Created: January 12, 2025
Last push: August 17, 2026

Star trust

Growth looks organic: forks and discussion are in line with active projects, and stars arrive unevenly, the way people give them.

These are heuristics, not a verdict: we judge by the repository’s behavior, not by a list of stargazers.

Spotted in

September 20, 2026GitHub Trending Python today: #5, +196 stars

README badge

Paste it into your README — the badge shows the current magnitude and links to this page.

[![Seismograph](https://gitnova.dev/badge/FareedKhan-dev/train-llm-from-scratch.svg?lang=en)](https://gitnova.dev/en/r/FareedKhan-dev/train-llm-from-scratch)

Similar projects

7.4
TianyuCodings/NanoJev
A nano replica of Jev built on Qwen3-0.6B that outputs probability distributions over candidates in a single forward pass with no token decoding, plus a training pipeline and game demos (maze, Snake).

Early signalLanguage modelsPython+325 stars in a day
7.2
mizorewww/laya-mlx
Native MLX runtime for Laya typed decision models: returns probabilities for choices, scores or truth values without text generation, running locally on Apple Silicon.

Early signalLanguage modelsPython+439 stars in a day
7.0
bespokelabsai/nimble
Nimble is a model and training recipe for fast typed decisions over text: given a flat schema of enum and boolean fields, it picks an answer and returns probabilities for each option. It targets routing, condition…

Early signalLanguage modelsPython+399 stars in a day
7.0
docling-project/docling
Docling is a document parsing library for many formats (PDF, DOCX, PPTX, XLSX, HTML, EPUB, audio, video) with advanced PDF understanding: page layout, tables, formulas, OCR. It prepares structured data for generative…

BreakoutLanguage modelsPython+334 stars in a day
6.9
githubnext/localjev
A local TypeScript/Bun HTTP service implementing a Jev-compatible POST /v1/systemone API on top of an OpenAI-compatible Chat Completions endpoint (DiffusionGemma via oMLX). It acts as a bridge for typed decisions…

Early signalLanguage modelsTypeScript+194 stars in a day
6.8
jaredpalmer/kev
kev is a LoRA adapter with a small readout head on top of Qwen2.5-0.5B that answers many typed questions about a document in a single forward pass, returning calibrated probabilities instead of text.

Early signalLanguage modelsPython+246 stars in a day