waxal-lin-omniasr-llm-3b

OmniASR LLM-3B v2, fine-tuned for Lingala as part of the WAXAL ASR solution.

Released artifact: top-3 FP64 parameter average. Decoded with beam 5, length_norm=True.

Files

File	Bytes
model.pt	17,522,653,578
omniASR_tokenizer_written_v2.model	91,481
card.yaml	—
requirements.txt	—

Total weights: 17,522,653,578 bytes (17.52 GB).

Role in the pipeline

Member of the Lingala TTIA fusion.

This model is one component of an ensemble solution and is not intended to be used alone. The routing, decoding, fusion and post-processing pipeline are in the solution repository.

How to use

This is a fairseq2 / omnilingual-asr checkpoint, loaded through the asset card shipped alongside it (card.yaml). The solution repository wraps the whole contract -- pinned environment, batch-size-1 decode, tokenizer wiring -- in one CLI:

git clone https://github.com/DariusTheGeek/waxal-asr-solution
cd waxal-asr-solution && bash install.sh        # pinned environments, ~15 min
python models/download_models.py --repo waxal-lin-omniasr-llm-3b

.venvs/omni/bin/python inference/decode/omniasr.py \
    --config configs/lin/llm3b.yaml \
    --audio path/to/wav_dir --output transcripts.csv

--weights accepts any directory holding this repo's files, e.g. the path returned by huggingface_hub.snapshot_download("DariusTheGeek/waxal-lin-omniasr-llm-3b"). requirements.txt in this repo pins the runtime alone; the environment lock the release was verified under is env/requirements-omni.txt in the solution repository.

Provenance

Parent	facebook/omniASR-LLM-3B-v2
Language	Lingala
Fine-tuning data	Waxal Lingala/Shona supervised split (google/WaxalNLP)
Seed	42

Usage

See https://github.com/DariusTheGeek/waxal-asr-solution for the environment locks, decode configuration and the exact command that reproduces the submission end to end from audio.

Licence

apache-2.0, inherited from the parent model.