gopt is an attempt at building a small Go code model for learning.

• builds a Go-focused corpus from public repos
• trains a BPE tokenizer for Go identifiers and literals
• trains a small decoder-only transformer
• samples code from prompts
• evaluates outputs with gofmt and go build

See SETUP.md for the actual workflow and commands.

• not a general LLM framework
• not a polished product
• not a frontier coding model

Because the interesting part is seeing the whole system work:

• corpus design
• tokenizer design
• dataset compilation
• model training
• sampling
• evaluation

The repo is intentionally small enough to understand and modify.