SmolLM2-135M – From Scratch
This project is a from-scratch PyTorch implementation of SmolLM2-135M.
What this demonstrates
- Architecture reverse-engineered (RMSNorm, RoPE, GQA)
- Numerical match with Hugging Face reference model
- Fine-tuned on TinyStories-Instruct
- FP16 safetensors export
- Deployed via Hugging Face Spaces
Usage
Enter a prompt and generate text using greedy decoding.
Notes
This is an educational + research project demonstrating full LLM reproduction.
