Youssefx64/Qareed-Qwen2.5-1.5B-Poet-v1
Youssefx64/Qareed-Qwen2.5-1.5B-Poet-v1 is a 1.5 billion parameter QLoRA fine-tuned model based on Qwen/Qwen2.5-1.5B-Instruct, developed by Youssefx64. This model is specifically designed for generating Arabic poetry, focusing on producing verses in the sadr/ajz format. It was trained on the arbml/Ashaar_dataset, with a context length of 32768 tokens, making it suitable for creative text generation tasks in Arabic. Its primary strength lies in following structural requests for poetry, though poetic quality and thematic adherence may vary.
Loading preview...
Overview
Youssefx64/Qareed-Qwen2.5-1.5B-Poet-v1 is an experimental 1.5 billion parameter model, fine-tuned using QLoRA from the Qwen/Qwen2.5-1.5B-Instruct base. Its core purpose is the generation of Arabic poetry, specifically trained to produce verses in the traditional sadr/ajz format. This model represents an early release (v1) in its development.
Key Capabilities
- Arabic Poetry Generation: Specializes in creating poetic lines in Arabic.
- Structural Adherence: Capable of following requests for a specific number of "baits" (couplets) and maintaining the sadr/ajz structure.
- Direct Output: Provides poetic output without additional explanations or conversational text.
Training Details
The model was trained using QLoRA (NF4) with r=16 and alpha=32 across all linear layers, for one epoch on approximately 20,000 instruction samples. The training data utilized was the arbml/Ashaar_dataset, which is a cleaned and deduplicated collection of poem-level samples, with colloquial poems removed. Diacritics were stripped from the training data, and meter (wazn) was not used as a control signal in this version.
Known Limitations
- Poetic Quality: Rhyme consistency and meter control are limited, and thematic adherence can be inconsistent.
- Textual Anomalies: Some words may be incorrect due to the loss of the alef-with-hamza-below character in the source text.
- Data Noise: Theme labels from the source websites are noisy, potentially affecting thematic generation.
- Reproduction: May reproduce lines directly from its training data, requiring human review of outputs.