WhisperKitPro: local model dir without Hugging Face defaults

[arda-argmax]

Summary

Adds a whisperkitpro-local-model pipeline alias for running WhisperKitPro against a local model directory only (no default Hugging Face repo_id / model_variant), and extends WhisperKitProConfig so model_dir alone selects --model-path without downloading.

Changes

• WhisperKitProConfig / WhisperKitPro
  • model_dir only: if set to a non-empty path, require an existing directory and use it for --model-path; no snapshot_download.
  • HF path: unchanged when model_dir is unset and both repo_id and model_variant are set (non-empty).
  • Legacy mode: unchanged; validation now errors if none of local model_dir, HF pair, or legacy triple is valid.
  • Treat empty strings like missing values via _config_str_provided.
• Pipeline alias whisperkitpro-local-model
  • Defaults: model_dir from WHISPERKITPRO_LOCAL_MODEL_PATH, cli_path from WHISPERKITPRO_CLI_PATH.
• Docs / config copy
  • WhisperKitProTranscriptionConfig and WhisperKitProConfig field descriptions updated to match the three modes (local / HF / legacy).

How to run

export WHISPERKITPRO_CLI_PATH=/path/to/whisperkitpro-cli
export WHISPERKITPRO_LOCAL_MODEL_PATH=/path/to/existing/model/dir

LOCAL_DATASET_PATH=/path/to/dataset LOCAL_DATASET_SPLIT=test \
uv run openbench-cli evaluate \
  --pipeline whisperkitpro-local-model \
  --dataset local-dataset \
  --metrics wer keyword_fscore keyword_precision keyword_recall

Breaking / behavior note

If model_dir is set but does not exist, the run fails with FileNotFoundError; there is no fallback to Hugging Face (even when repo_id / model_variant are also set).

[EduardoPach]

lgtm!

[dbrkn]

lgtm