Downloads · 30 days
6
10% of all-time downloads
CuriousDragon/ettin-17m-sponsor-v3-replay
ettin-17m-sponsor-v3-replay is a token classification model from CuriousDragon. Use it when you need labels on individual words, such as names. It is set up for transformers. The card lists the license as cc-by-nc-sa-4.0.
Release-candidate weights. Initialized from v1 and trained for two epochs on 12,048 corrected/replay windows at 1e-6. Improves precision on self-promo and similar negatives at some recall cost. Calibrated decoder thre…
Downloads · 30 days
6
10% of all-time downloads
All-time downloads
58
Public
Parameters
16.9M
67.5 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors67.5 MB · 95%
From the Hugging Face model README
Release-candidate weights. Initialized from v1 and trained for two epochs on 12,048 corrected/replay windows at 1e-6. Improves precision on self-promo and similar negatives at some recall cost. Calibrated decoder threshold is 0.65.
jhu-clsp/ettin-encoder-17m @ 987607455c61e7a5bbc85f7758e0512ea6d0ae4cModernBertForTokenClassification)O, B-SPONSOR, I-SPONSOR, L-SPONSOR, U-SPONSORCuriousDragon/ettin-17m-sponsor-v3-replayfrom transformers import AutoModelForTokenClassification, AutoTokenizer
repo = "CuriousDragon/ettin-17m-sponsor-v3-replay"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForTokenClassification.from_pretrained(repo)
For full YouTube transcripts, tokenize into overlapping 768-token windows
(128-token overlap), decode BILOU spans, map characters back onto caption
cues, and merge gaps of 24 normalized characters or 1500 ms. See the Flow
ml/sponsor_detection inference pipeline.
| Split | Sponsor token P / R / F1 | Token accuracy |
|---|---|---|
| Validation | 0.933 / 0.898 / 0.915 | 0.974 |
| Test | 0.937 / 0.886 / 0.911 | 0.972 |
Validation, character IoU 0.5: P 0.935 / R 0.846 / F1 0.888
Held-out test windows:
On-device or local inference over English video transcripts to propose paid sponsor ranges. In Flow, SponsorBlock remains authoritative when it has segments; these weights are a fallback / shadow detector.
decoder.json is bound to these weights by SHA-256. Do not mix it with another checkpoint.
Related: CuriousDragon/ettin-17m-sponsor-v1
These weights are released under CC BY-NC-SA 4.0.
The base encoder jhu-clsp/ettin-encoder-17m
is MIT. Fine-tuning used SponsorBlock labels
(CC BY-NC-SA 4.0) and the pinned Xenova/sponsorblock-768
window corpus. Non-commercial use and share-alike apply to this derivative.
This project is not affiliated with SponsorBlock.