CuriousDragon/flow-smart-segments-20261004

Flow Smart Segments — 2026-10-04

English transcript passage detection for paid sponsorship (sponsor), unpaid/self-promotion (selfpromo), and generic engagement requests (interaction).

This is the finalized locally qualified Flow model, version curated-category-encoder-ensemble-20261004. It has two Ettin encoders: one completed sponsor branch, and one completed branch shared by self-promotion and interaction. Each category independently predicts BILOU tags. This custom architecture is not compatible with the standard Transformers AutoModel token-classification loader.

Evaluation

Actual full INT8 prediction path on 159 fixed, AI-reviewed development windows:

Category	Matched passages	Reference recall	Weak-absence windows with detections
Sponsor	16/17	94.1%	1/130 (0.8%)
Self-promotion	7/12	58.3%	7/145 (4.8%)
Interaction	12/23	52.2%	6/139 (4.3%)

All predefined local gates passed; all 14 original sponsor anchors were retained. FP32 exported-graph matches were 16/17, 7/12 and 13/23; INT8 exported-graph matches were 16/17, 7/12 and 12/23.

These are development reference matches at character IoU >= .5, not human-verified accuracy, precision or F1. Weak-absence detections are alarms on incomplete supervision, not confirmed false positives. Self-promotion and interaction still miss substantial numbers of passages. See evaluation.json for the actual summarized receipts and their source hashes.

Training lineage

Both branches descend from the same fresh-base fit on 370,488 eligible transcript windows, 2 complete epochs / 740,976 exposures. The sponsor branch then completed one epoch over 1,535 reviewed windows. The SP/I branch completed 16 epochs over all 2,125 usable reviewed windows, including a new append-only 600-window source addendum. These reviewed sets overlap and must not be added together. Assembly copied completed parent weights without averaging, reinitialization or another optimizer fit. Development/FINAL text was not used for training.

Base: jhu-clsp/ettin-encoder-17m, revision 987607455c61e7a5bbc85f7758e0512ea6d0ae4c. Training uses SponsorBlock source labels, Xenova/sponsorblock-768, and ScriptSmith/sponsorblock-youtube-metadata-2024, with incomplete supervision kept UNKNOWN and provisional source-only AI reviews.

Files and use

• Root model.safetensors + config.json: full custom Torch checkpoint.
• runtime/: byte-exact installed production bundle, with tokenizer, decoder, manifests, rank-4 FP32/INT8 ONNX graphs and rank-3 category exports.
• code/: matching Python runtime implementation.
• publication_manifest.json: SHA-256 pins for every uploaded payload file.
• requirements.txt: versions used by the verified local environment.

Download and run the bundled public INT8 predictor:

from pathlib import Path
import sys
from huggingface_hub import snapshot_download

download = Path(snapshot_download("CuriousDragon/flow-smart-segments-20261004"))
sys.path[:0] = [str(download / "code"), str(download / "code/src")]
from scripts.predict_curated_production import predict
from sponsor_detection.inference.windowing import TranscriptCue

result = predict(
    download / "runtime",
    [TranscriptCue(0, 0, 1, "This episode is brought to you by our sponsor.")],
    timestamps=False,
    precision="int8",
)
print(result["spans"])

For real caption cues, supply ordered millisecond start/end times and timestamps=True. Times within cues are interpolated. The predictor uses 1024-token windows with 128-token overlap. Thresholds are sponsor .7 / selfpromo .875 / interaction .825; the minimum passage length is 12 characters. The final decoder preserves distinct events within each window and deduplicates overlapping windows, with zero additional character-gap merging.

For Torch loading, use code/src/sponsor_detection/model/multi_head_model.py::load_saved_multi_head(download, revision="987607455c61e7a5bbc85f7758e0512ea6d0ae4c"). The output has shape [batch, sequence, 3, 5], category order sponsor/selfpromo/interaction and tags O/B/I/L/U.

Limitations and license

English only. This release has not been independently validated on unseen full videos or deployed in Android. The rank-4 interface is not the legacy Android interface; rank-3 exports are also undeployed. Automatic playback actions remain disabled. No transcripts, datasets, optimizer states or authentication secrets are included in this model upload.

These derivative weights follow this project's CC BY-NC-SA 4.0 release policy. The base encoder is MIT; SponsorBlock labels are CC BY-NC-SA 4.0. ScriptSmith's compilation is CC BY 4.0, while source YouTube captions remain subject to their original terms. This project is not affiliated with SponsorBlock.

Code repository: https://github.com/seshuthota/Flow-SponsorML