Single Go application with:

• Go backend API
• Go-served static frontend (no Node.js, no Vite, no TypeScript)
• Small educational character-level transformer/autograd implementation

Important scope note:

• This is a toy model trained from scratch on user-provided strings at runtime.
• It is not a pretrained production GPT model.

• Go 1.21+

go run .

Open:

• http://127.0.0.1:8080/
• Docs page: http://127.0.0.1:8080/docs/

Live website:

• http://atomicgpt.duckdns.org:8080/

go build ./...

• main.go: app bootstrap (embed assets, register routes, start server)
• server.go: HTTP handlers and shared server state
• api_types.go: request/response structs for API
• autograd.go: tiny autodiff engine (Value, ops, Backward)
• model.go: model config/state, initialization, math helpers, optimizer
• forward.go: transformer forward pass
• inference_and_training.go: training step + sampling + trace generation
• web/index.html: main UI
• web/app.js: browser logic
• web/docs/index.html: help page

1. POST /api/init

• Purpose: initialize model with documents and hyperparameters.
• Body:

{
  "docs": ["alex", "anna", "john"],
  "config": {
    "n_embd": 16,
    "n_head": 4,
    "n_layer": 1,
    "block_size": 16,
    "learning_rate": 0.05
  }
}

2. POST /api/train

• Purpose: train model parameters.
• Body is optional.
• Optional body fields:

{
  "steps_per_call": 1,
  "batch_size": 6
}

• Defaults when omitted:
• steps_per_call = 1
• batch_size = 6

3. POST /api/generate

• Purpose: sample generated text.
• Body is optional.
• Optional body fields:

{
  "options": {
    "temperature": 0.7,
    "top_k": 5,
    "min_len": 3
  }
}

• Defaults when omitted:
• temperature = 0.7
• top_k = 5
• min_len = 3

4. POST /api/generate_trace

• Purpose: sample generated text and return per-step sampling trace.
• Accepts the same optional options as /api/generate.