title | emoji | colorFrom | colorTo | sdk | app_file | pinned |
---|---|---|---|---|---|---|
Sponsorblock ML |
🤖 |
yellow |
indigo |
streamlit |
app.py |
true |
Automatically detect in-video YouTube sponsorships, self/unpaid promotions, and interaction reminders. The model was trained using the SponsorBlock database licensed used under CC BY-NC-SA 4.0.
Check out the online demo application at https://huggingface.co/spaces/Xenova/sponsorblock-ml, or follow the instructions below to run it locally.
-
Download the repository:
git clone https://github.com/xenova/sponsorblock-ml.git cd sponsorblock-ml
-
Install the necessary dependencies:
pip install -r requirements.txt
-
Run the application:
streamlit run app.py
-
Predict for a single video using the
--video_id
argument. For example:python src/predict.py --video_id zo_uoFI1WXM
-
Predict for multiple videos using the
--video_ids
argument. For example:python src/predict.py --video_ids IgF3OX8nT0w ao2Jfm35XeE
-
Predict for a whole channel using the
--channel_id
argument. For example:python src/predict.py --channel_id UCHnyfMqiRRG1u-2MsSQLbXA
Note that on the first run, the program will download the necessary models (which may take some time).
This is primarly used to measure the accuracy (and other metrics) of the model (defaults to Xenova/sponsorblock-small).
python src/evaluate.py
In addition to the calculated metrics, missing and incorrect segments are output, allowing for improvements to be made to the database:
- Missing segments: Segments which were predicted by the model, but are not in the database.
- Incorrect segments: Segments which are in the database, but the model did not predict (meaning that the model thinks those segments are incorrect).
This can also be used to moderate parts of the database. To moderate the whole database, first run:
python src/preprocess.py --do_process_database --processed_database whole_database.json --min_votes -1 --min_views 0 --min_date 01/01/2000 --max_date 01/01/9999 --keep_duplicate_segments
followed by
python src/evaluate.py --processed_file data/whole_database.json
The --video_ids
and --channel_id
arguments can also be used here. Remember to keep your database and processed database file up-to-date before running evaluations.
-
Download the SponsorBlock database
python src/preprocess.py --update_database
-
Preprocess the database and generate training, testing and validation data
python src/preprocess.py --do_transcribe --do_create --do_generate --do_split --model_name_or_path Xenova/sponsorblock-small
--do_transcribe
- Downloads and parses the transcripts from YouTube.--do_create
- Process the database (removing unwanted and duplicate segments) and create the labelled dataset.--do_generate
- Using the downloaded transcripts and labelled segment data, extract positive (sponsors, unpaid/self-promos and interaction reminders) and negative (normal video content) text segments and create large lists of input and target texts.--do_split
- Using the generated positive and negative segments, split them into training, validation and testing sets (according to the specified ratios).
Each of the above steps can be run independently (as separate commands, e.g.
python src/preprocess.py --do_transcribe
), but should be performed in order.For more advanced preprocessing options, run
python src/preprocess.py --help
The transformer is used to extract relevent segments from the transcript and apply a preliminary classification to the extracted text. To start finetuning from the current checkpoint, run:
python src/train.py --model_name_or_path Xenova/sponsorblock-small
If you wish to finetune an original transformer model, use one of the supported models (t5-small, t5-base, t5-large, t5-3b, t5-11b, google/t5-v1_1-small, google/t5-v1_1-base, google/t5-v1_1-large, google/t5-v1_1-xl, google/t5-v1_1-xxl) as the --model_name_or_path
. For more information, check out the relevant documentation (t5 or t5v1.1).
The classifier is used to add probabilities to the category predictions. Train the classifier using:
python src/train.py --train_classifier --skip_train_transformer