CLI reference
Every command `zrodlo` runs, each with its own page: what it does, every flag, and what it refuses.
How commands read#
Commands are named object then verb: tokenizer, corpus, model and checkpoint are the objects, and commands list describes them all. Every number comes from the command line or config.json; Zrodlo chooses no size, rate, decay, epsilon, probability or seed. A result prints on stdout as field: value lines, nested fields joined by dots, or as one JSON document with --json. Progress prints on stderr. A refusal prints Error: and its cause on stderr and exits with status 1; a missing or malformed flag is a usage error and exits with status 2. A command never overwrites an existing output, and a delete removes nothing without --confirm.
Commands#
| Command | What it does |
|---|---|
| tokenizer train | Train a byte-level BPE tokenizer on text files and write it as a new tokenizer.json |
| tokenizer inspect | Load a tokenizer and report its vocabulary size, special tokens and digest |
| tokenizer delete | Remove a tokenizer file; refused without --confirm |
| corpus prepare | Encode text and JSON Lines files with the model's tokenizer into tokens.bin and corpus.json in a new directory |
| corpus inspect | Validate a corpus against its manifest and report the manifest |
| corpus split | Write a corpus's last documents to a new held-out corpus and the rest to a new training corpus, without encoding them again |
| corpus delete | Remove a corpus directory; refused without --confirm |
| model init | Write a freshly initialized model directory from a configuration |
| model train | Train every weight of a model directory, from scratch or continued |
| model evaluate | Score a model directory's saved weights on a corpus; no weight changes |
| model generate | Continue a prompt with a model's saved weights, greedily or by a seeded draw; no weight changes |
| model inspect | Validate a model directory or checkpoint and summarize it |
| model delete | Remove a model directory or checkpoint; refused without --confirm |
| checkpoint list | List a run's step-<n> checkpoints with their steps and last loss |
| checkpoint resume | Continue a checkpoint with its own optimizer state and data position |
| commands list | Every command with its flags, whether each is required, and the values it accepts |