SmolGuru-80M-512

This public repo is being prepared for an ~80M parameter LLaMA-style chatbot trained from scratch with the HuggingFaceTB/SmolLM2-135M-Instruct tokenizer.

Status: scaffold/config repo created. Trained weights will be uploaded after pretraining + SFT complete.

Planned training:

• 2B high-grade pretraining tokens
• 75k filtered SFT examples
• 512 context
• 256-token answer target
• Final + best checkpoint retained