ster toy-model
Write a tiny offline Llama-family checkpoint that Ster can load, so every command can be exercised end to end on a laptop with no download, no GPU and no account.
Invocation and purpose#
ster toy-model <DIR>The checkpoint is real in shape only: config.json says model_type: "llama" with the smallest grouped-query shape that exercises every path Ster's decoder has (four layers, four heads over two key-value heads, hidden size 64), tokenizer.json is a WordLevel tokenizer over a vocabulary of about sixty words, and model.safetensors holds seeded random weights in the exact tensor layout Ster loads. Generated text is deterministic gibberish drawn from the toy vocabulary; the point is the mechanics, not the prose. The same seed on every run, so two toy checkpoints are byte-for-byte identical.
Inputs and options#
| Input | Required | Meaning |
|---|---|---|
<DIR> | yes | The directory the checkpoint is written into; created, with its parents, when it does not exist. |
Output and state#
Writes config.json, tokenizer.json and model.safetensors into the directory, replacing those three files when they are already there, and prints the directory path. The directory is then a --model for every other command.
Refusals#
create <path>, with the cause, when the directory cannot be created.- A file that cannot be written is refused with the filesystem's own error and the path.