Someone squeezed pretraining, SFT, reward modeling, DPO, PPO and GRPO into a…
This is a dev post classified by Jev as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
Someone squeezed pretraining, SFT, reward modeling, DPO, PPO and GRPO into a 13M-parameter LLM you can run on a free T4. The repo builds the Transformer in plain PyTorch, then carries the same model through the full training path. The useful part is that the loops are actually there to read, not buried inside transformers or trl, so you can see exactly what changes at each stage. One of the cleanest LLM training repos I’ve seen. https://github.com/FareedKhan-dev/train-llm-from-scratch
Posted by Sandhya (1.3k followers) 58 min ago · 0 likes · 20 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: train-llm-from-scratch.
More dev work like this
- Stepfun just announced their new SOTA model — @notjazii
- Coding agents get expensive when every task gets premium-model treatment. — @DanKornas
- iMessage was never meant to win. — @xadisingh
- Welcome back, StepFun! — @Hikari_07_jp
- The lines between text and image gen models are blurred a bit. — @loktar00
- lmaooo https://classifier.dev already has copycats — @michael_chomsky
- Your AI agent knows your product. This repo gives it an SEO workflow. — @DanKornas
- Building a Recurrent Neural Network. #BigData #Analytics #DataScience #AI… — @gp_pulipaka
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 15.3k posts from 4.8k X accounts over the last 21 days, 1.7k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-20 07:24 UTC. Full method.