gonnerthetooner/orislop-av-joint

Orislop AV Joint

Orislop AV Joint is an experimental, multitask audio-visual forensics training kit. It contains a T4-oriented Google Colab workflow, a PyTorch model trained from scratch, bounded AV preprocessing, evaluation/calibration/export commands, and a rights-aware downloader for approved open audiovisual sources.

Repository status

This repository contains training code, not trained or production-approved weights. Do not present it as a working universal deepfake detector until a checkpoint has passed the documented data, evaluation, calibration, fusion, and shadow-mode gates.

The exported artifact remains unpromoted by default. Standalone AV output is not authorized to hide or skip content.

Colab quick start

1. Download this repository as a ZIP.
2. Open Orislop_AV_Joint_Open_Data_Trainer.ipynb in Google Colab.
3. Select a T4 GPU.
4. Run the notebook in order.
5. To acquire AMI data, read the linked AMI license and consent evidence, then change the two explicit acceptance switches in the acquisition cell.

The acquisition tool is not a general web scraper. It only follows the official AMI corpus host allowlist, caps completed AV media at 50 decimal GB, resumes partial files, rejects off-host redirects, and records SHA-256 provenance.

On Windows, tools/run_local_data_pipeline.ps1 provides plan, discovery, download, manifest-build, and full modes. Discovery accepts no license and downloads no media. Full mode requires the explicit -AcceptAmiLicense switch, checks target-drive capacity, and defaults to D:\OrislopAVData.

If the standard Hugging Face CLI reports a Windows certificate-chain error, tools/hf_native_tls.py runs the same CLI through the operating system trust store. tools/hf_native_tls_token_login.py provides a hidden-input token login without putting the credential in a command line.

The paired manifest generator reads AMI's official meeting-specific camera/headset mapping. It produces both synchronized examples and controlled audio offsets. Controlled offsets are labeled as legitimate delay (sync_mismatch=1, joint_forgery=0), not synthetic-media fraud.

Open-source and data boundary

Code, trained weights, and datasets have separate licenses. Read:

• OPEN_SOURCE_PATH.md
• data_sources.json
• AMI_ATTRIBUTION.md
• DATASET_LABELING_GUIDE.md

No project-level software license has been asserted in this model card. Add a reviewed LICENSE before claiming that the repository or future weights are fully open source.

AMI-only training is useful for authentic AV representation and synchronization experiments, but it is not sufficient for a general deepfake detector. Such a release also needs rights-cleared manipulated examples, legitimate dubbing and delay negatives, held-out generator families, speaker/source-disjoint splits, calibration, independent evaluation, and reviewed shadow decisions.

Safety and limitations

• Predictions are probabilistic signals, not proof of deception.
• Ordinary dubbing, translation, editing delay, network lag, and accessibility audio must not be labeled as forgery solely because AV timing differs.
• The detector must abstain when face coverage, speech, motion, SNR, or uncertainty gates fail.
• Do not use it to make employment, credit, housing, identity, or legal decisions.
• Do not redistribute source media unless its license and consent terms permit that exact use.
• Treat synchronization mismatch as a forensic signal, never as proof that a person lied or that a video is malicious.