llamersWatch a language model
actually think.
llamers is a complete, from-scratch transformer — tokenizer, attention, RoPE, KV cache, sampling — running live in pure JavaScript. No mocks, no magic: every number on screen comes out of a real trained weight. Open the hood and watch every token take shape.
- See the attention — Per-head heatmaps update as each token is generated.
- Watch the cache grow — The KV cache fills one position at a time — exactly as in real runtimes.
- Step into the math — Walk one operation at a time through a single token's forward pass.
Now feed one real data.
Dataset Studio is the companion tool: an IDE for LLM training datasets. Edit .jsonl records, validate them against the model's own chat template — the layer where reasoning traces get silently dropped and misspelled content parts vanish — and preview the exact prompt text with the loss mask your training run will use.