ster decisions import
Read a JSONL file of context, options and label rows into labelled decisions, so an option-scoring set built elsewhere trains and benchmarks here without a converter.
Invocation and purpose#
ster decisions import --input <JSONL> --question-id <ID> --instructions <TEXT> --output <OUTPUT>Each line is one JSON object with context (the text), options (a list of texts) and label (the index of the correct option). The context becomes the example's state, the options become a choice question's options in the row's order, and the label index names the answer. The question's id and wording are the caller's: the file carries neither, and none is assumed.
Inputs and options#
| Input | Required | Meaning |
|---|---|---|
--input <JSONL> | yes | The file, one row per line; blank lines are ignored. |
--question-id <ID> | yes | The id the one question is asked under in every example. |
--instructions <TEXT> | yes | The question asked about every context. |
--output <OUTPUT> | yes | Where the labelled decisions are written; missing parent directories are created. |
Output and state#
Prints an object with path and report; the report carries source, rows read, examples kept and rejected, one sentence per refused row with its line number: line <n>: not a context/options/label row: <error>, the context is empty, a row needs at least two options, a row has <n> options; Ster labels at most 26, an option is empty, two options carry the same text, or the label <i> names no option of <n>. A rejected row is left out and the rest are kept.
Refusals#
- A missing
--question-idor--instructionsis refused by clap with exit 2. an imported question needs an id;an imported question needs instructionswhen either is blank.failed to read <path>;<path> holds no importable rowswhen every row was rejected; the labelled-decision validation refusals of `ster calibrate`.