Downloads · 30 days
4
6% of all-time downloads
Abdoul27/waxal-af51-linsna-ck1200
waxal-af51-linsna-ck1200 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. The card lists the license as mit.
Full fine-tune (1.54B params) on Google WaxalNLP, Lingala (ln 50353) and Shona (sn 50324). 31,545 clips, speaker-disjoint holdout (20 speakers, 500 eval clips, zero overlap). RAW cased labels; normalisation only at sc…
Downloads · 30 days
4
6% of all-time downloads
All-time downloads
67
Public
Parameters
1.5B
3.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3.1 GB · 100%
From the Hugging Face model README
Full fine-tune (1.54B params) on Google WaxalNLP, Lingala (ln 50353) and
Shona (sn 50324). 31,545 clips, speaker-disjoint holdout (20 speakers,
500 eval clips, zero overlap). RAW cased labels; normalisation only at scoring.
lr 1e-5 linear / 500 warmup / batch 8 x accum 4 / bf16 / grad checkpointing.
| step | score | lin | sna |
|---|---|---|---|
| 0 (zero-shot) | 0.8854 | W0.170/C0.068 | W0.192/C0.028 |
| 300 | 0.8795 | W0.188/C0.080 | W0.187/C0.027 |
| 600 | 0.8787 | W0.185/C0.076 | W0.196/C0.028 |
Validation declines from step 0. Published to test whether that reflects real behaviour, or whether WAXAL-derived validation is unreliable for a base model that may already have seen the corpus.
Force the language at decode: forced_decoder_ids=[(1, 50353), (2, 50360), (3, 50364)]
for Lingala, 50324 for Shona.