youtube-laya-multilingual

Fine-tune of convaiinnovations/laya-multilingual (Apache-2.0) for the YouTube Community Guidelines. The base model is mmBERT, vocab 256k. Laya scores text: a title, description, comment, transcript, or policy excerpt. It does not watch a video.

Trained with RLCD on the Community Guidelines in the 80 Help Center languages that are real translations and that mmBERT can read. A later pass adds short ordinary video titles and short violation lines. Guideline sections alone are long, and the short ones are mostly violations, so everyday titles were scoring as not allowed. The guideline hold-out and the four external benches below were not part of training. Noul temperature is 0.5. Threshold is 0.5.

Positive class: content the guidelines do not allow. Metric: Binary F1. The same not_allowed question is used for the zero-shot baseline and this checkpoint. Pass the raw text. A JSON object is a different string from the one used in training and in the demo.

Accuracy

The pre-registered bar is macro accuracy on the guideline hold-out: the unweighted mean of per-language accuracies. External benches are transfer checks and are not part of that bar. Both columns below are scored on the raw text.

Model	guideline hold-out	YouToxic toxic	YouToxic hate	YouToxic obscene	YouToxic threat	UCI YouTube spam	Jigsaw multilingual	Multilingual HateCheck
laya-multilingual (zero-shot)	0.850	0.664	0.406	0.418	0.123	0.619	0.489	0.629
youtube-laya-multilingual (this checkpoint)	0.969	0.708	0.401	0.369	0.091	0.743	0.561	0.695

WIN — hold-out macro accuracy 0.9487 beats zero-shot 0.7409. Hold-out Binary F1 is 0.969 (precision 0.950, recall 0.989) against 0.850 (precision 0.770, recall 0.948). Policy-bucket accuracy is 0.797 against 0.206.

YouToxic, UCI spam, Jigsaw, and HateCheck move by small amounts, except spam and Jigsaw, which rise. They cover spam, toxicity, and hate. Misinformation, EDSA, child safety, and regulated goods stay on the guideline hold-out.

Inference speed

Wall-clock per sample on the guideline hold-out (452 sections). Jigsaw timing was not recorded.

Model	ms/sample
laya-multilingual (zero-shot)	3.6
youtube-laya-multilingual (this checkpoint)	3.5

Usage

import laya

agent = laya.load("3p3r/youtube-laya-multilingual", device="cuda")
questions = {
    "not_allowed": {
        "type": "noul",
        "instructions": "Does the following text describe content that YouTube's Community Guidelines do not allow?",
        "criteria": {
            "false": "content the guidelines allow, including an EDSA exception",
            "true": "content the guidelines do not allow",
        },
    }
}
result = agent.predict("Buy 10000 subscribers. WhatsApp me.", questions)
print(result["answers"]["not_allowed"]["noul"] >= 0.5)

Citation

Cite this model:

@misc{youtube-laya-multilingual2026,
  title        = {youtube-laya-multilingual: YouTube Community Guidelines on multilingual Laya},
  author       = {{3p3r}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/3p3r/youtube-laya-multilingual},
  note         = {Fine-tune of \url{https://huggingface.co/convaiinnovations/laya-multilingual}}
}