Downloads · 30 days
2
3% of all-time downloads
Abdoul27/waxal-mms1b-lin-v4-ck3000
waxal-mms1b-lin-v4-ck3000 is a machine learning model from Abdoul27. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
facebook/mms-1b-all with the pretrained lin adapter and CTC head, fine-tuned on WAXAL Lingala. Fresh single linear schedule (not a resume), batch 8 x accum 4 = 32 effective, lr 3e-5 body / 3e-4 head, warmup 400, 8 spe…
Downloads · 30 days
2
3% of all-time downloads
All-time downloads
76
Public
Parameters
965M
3.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3.9 GB · 100%
From the Hugging Face model README
facebook/mms-1b-all with the pretrained lin adapter and CTC head,
fine-tuned on WAXAL Lingala. Fresh single linear schedule (not a resume),
batch 8 x accum 4 = 32 effective, lr 3e-5 body / 3e-4 head, warmup 400,
8 speakers held out.
Zero-shot baseline on the speaker-disjoint holdout: 0.8438 (WER 0.2301 / CER 0.0823).
proc = AutoProcessor.from_pretrained(REPO) # target_lang already set to lin
model = Wav2Vec2ForCTC.from_pretrained(REPO)
The v3 lineage decoded with an AfriVoice 5-gram KenLM at alpha 0.3 / beta 0.5, beam 100, worth +0.0079 on the Zindi Lingala probe.
Reference points: v3 (~4 epochs, +KenLM) probes 0.289448; the best submission combining it with AF51 ck1200 on Shona scores 0.749367. Screen checkpoints by probe, not by the local eval curve -- that curve has pointed the wrong way three times in this competition.