You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to this model is manually gated. Please describe your intended use.
Log in or Sign Up to review the conditions and access this model content.
sl-linnet
Chatterbox Multilingual v3 fine-tuned on Tajik (t3-only; s3gen/ve/conds are the frozen
Chatterbox components, grapheme_mtl_merged_expanded_v1.json is the same extended tokenizer used
by the lab's other Chatterbox-MTL Tajik checkpoints).
This is the control arm of a 3-point data-scaling comparison — identical training recipe
(lr 1e-5, fp32, batch 8 x grad-accum 2, warmup 300, language_id=ru, same pretrained warm-start)
to the previously shipped Tohirju/chatterbox-mtl-tajik (cb_step8000, trained on 205.5h /
100,000 utts), scaled up to 493h / 226,441 utts (tg_subset_full). Checkpoint step 12000
of 12,500 trained, selected by round-trip WER on a frozen n=60 Tajik eval set (never by loss).
Eval (frozen n=60 FLEURS-tg round-trip WER, identical set + scorer as cb_step8000)
| system | n | mean WER | median WER |
|---|---|---|---|
| this checkpoint (493h) | 60 | 9.30% | 6.38% |
cb_step8000 (205.5h, previously shipped) |
60 | 10.73% | 8.33% |
Delta: -1.43pp mean / -1.95pp median (this checkpoint better), well above the lab's ~0.2pp
seed-to-seed noise floor. Per-item on the identical 60-utterance set: 21/60 better, 25/60 tied,
14/60 worse. Raw per-utterance scores: eval/per_utt_n60.jsonl, eval/scores_n60.json.
Status: research/data-scaling result. A candidate to replace cb_step8000 as the product
default, pending a sibling Arm A run (662h Khazina corpus, still training as of this push) and
Tohir's own review — not yet promoted to product.
- Downloads last month
- -