whispered-mla-experiments

Whispered MLA — experiments zoo

Every study checkpoint behind the Whispered MLA release (specialists, mixes, fairness controls, plus_cross variants, matched-budget pairs), so the study can be verified or extended without retraining. Research artifacts — read the warnings.

Load any entry: AutoModelForSpeechSeq2Seq.from_pretrained('burakaydinofficial/whispered-mla-experiments', subfolder='<name>', trust_remote_code=True) (transformers==4.46.x; scheme=none entries load without trust_remote_code).

Warnings: single-language slate/* models DESTROY other languages (that finding is their purpose); *_tiny/_base are 8k-budget concept runs, superseded by the released family; quality claims live in the released models' cards, not here.

Full per-model provenance: registry.json in this repo.