Bayan Arabic Summarization Model

This repository contains the summarization checkpoint used by Bayan.

Recommended generation settings

Use the following defaults for the best balance of faithfulness and brevity:

summary = model.generate(
    **inputs,
    max_new_tokens=40,
    num_beams=1,
    do_sample=False,
    early_stopping=False,
    no_repeat_ngram_size=3,
    repetition_penalty=1.1,
)

Notes

• The checkpoint is optimized for Arabic news-style summarization.
• Beam search and sampling may increase hallucination on some inputs.
• If a summary drifts too far from the source, prefer a more extractive fallback.