kroma.pipeline

Matching run

MatchingRun is the template for one evaluation: prepare, score, then finish.

Call

MatchingRun(args, listener=None)
MatchingRun.execute() -> dict

execute calls prepare, score_pairs, and finish. listener defaults to MetricsListener, which counts tokens and refinement time.

Source: pipeline.py.

Methods

Method	What it does
`prepare()`	Load the method JSONL and dataset JSON, open output files, build the task, and fill the vector stores.
`score_pairs()`	Judge each alignment. Skip a pair already stored. Refine on a fixed interval.
`finish()`	Run the final offline refinement and write the metrics JSON.

Output

Accepted alignments go under results/. Pairs that need an expert go under reviews/. The metrics file sits next to the prediction file and ends in .metrics.json. It includes precision, recall, F1, token counts, and refinement timings.

The run log is logs/kroma.log.