metadata

title: Sponsorblock ML
emoji: 🤖
colorFrom: yellow
colorTo: indigo
sdk: streamlit
app_file: app.py
pinned: true

SponsorBlock-ML

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 demo application by visiting https://xenova.github.io/sponsorblock-ml/. You can also run it locally using make run.

Predicting

Download the repository:

git clone https://github.com/xenova/sponsorblock-ml.git
cd sponsorblock-ml

Run predictions:
- Predict for a single video using the --video_id argument. For example:
```
python src/predict --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).

Evaluating

Measuring Accuracy

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).

Moderation

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.

Training

Preprocessing

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
```
1. --do_transcribe - Downloads and parses the transcripts from YouTube.
2. --do_create - Process the database (removing unwanted and duplicate segments) and create the labelled dataset.
3. --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.
4. --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

Transformer

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).

Classifier

The classifier is used to add probabilities to the category predictions. Train the classifier using:

python src/train.py --do_train_classifier --skip_train_transformer