Qareed Poet 1.5B (v1) — قريض
Arabic poetry generation model, QLoRA fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. This is an early experimental release (v1).
Usage
from transformers import pipeline
pipe = pipeline("text-generation", model="Youssefx64/Qareed-Qwen2.5-1.5B-Poet-v1", device_map="auto")
msgs = [{"role": "user", "content": "اكتب بيتين من الشعر في الشوق."}]
print(pipe(msgs, max_new_tokens=120, do_sample=True, temperature=0.8)[0]["generated_text"][-1]["content"])
Training
- Data:
arbml/Ashaar_dataset(Ashaar, arXiv:2307.06218): cleaned, deduplicated, poem-level train/val split, colloquial poems removed - Method: QLoRA (NF4), r=16, alpha=32, all linear layers, 1 epoch, ~20k instruction samples (1-4 consecutive baits)
- Diacritics were stripped; meter was NOT used as a control signal in v1
- Final validation loss: about 1.337
What it does well
- Follows the requested number of baits and the sadr/ajz format; answers without explanations.
Known limitations
- Poetic quality is limited: rhyme is often inconsistent, meter (wazn) is not controlled or checked, and the requested theme is frequently not respected.
- Source text lost the letter alef-with-hamza-below; common words were patched, but some words may still be wrong.
- Theme labels come from the source websites and are noisy.
- May reproduce lines from the training data. Outputs need human review.