Ster/Docs

ster sources

List the data other tools keep that Ster reads, one adapter each, or say how many examples or pairs one file gives.

Invocation and purpose#

text
ster sources [--all]
ster sources discover <ADAPTER> --path <FILE>
ster sources share [yes|no]

Every source is an adapter with an id IMP-STER-<NN>. An example set another trainer published — OpenAI's fine-tuning JSONL (IMP-STER-03), ShareGPT conversations (04), Alpaca (05) — is read by `ster tune sft` --examples; preference rows — a Hugging Face preference dataset (01), TRL's DPO rows (06) — by `ster pairs import` --benchmark preference --source <file>, which writes a pair set. The pair set Preferences exports (19: trait_name and pairs of positive and negative) is already one, so `ster workspace import` or --pairs reads it as it is, and discover counts its pairs. The model checkpoints --model loads are listed beside them. Nothing is written: the writing verbs are those commands.

Inputs and options#

InputRequiredMeaning
--allnoList every adapter; Ster reads the same files on every system.
discover <ADAPTER> --path <FILE>—Read the file as that adapter's command would, and say how many examples or pairs it gives, or why it is not that source.
share [yes|no]—Whether each discovery reports its adapter and counts to Wisent Integrations; alone, it prints the question.

Output and state#

The list is a pretty JSON document with adapters (each with adapter, source, kind, read_by and found), all and the sharing question. A discovery answers that adapter's row with path, found, examples or pairs, and left_out (rows that give none), or why_not when the file is not that source.

Refusals#

  • An adapter Ster does not have: ster.sources.sources-adapter-unknown, naming the adapters there are (exit 2).
  • An answer to sources share that cannot be kept in the configuration directory: ster.sources.sources-share-unkept (exit 1); set XDG_CONFIG_HOME to a directory this user can write.