tdw419/judaism-llm-qwen2.5-7b
The tdw419/judaism-llm-qwen2.5-7b is a 7.6 billion parameter language model, fine-tuned from Qwen/Qwen2.5-7B-Instruct using LoRA adapters. It specializes in Jewish texts, including Torah, Talmud, Halakha, and commentaries, trained on 6,112 Sefaria texts. This model excels at generating structured, bilingual Hebrew/English responses with accurate Jewish knowledge, making it suitable for analysis of religious texts and related discourse.
Loading preview...
Overview
This model, tdw419/judaism-llm-qwen2.5-7b, is a 7.6 billion parameter language model fine-tuned from the Qwen/Qwen2.5-7B-Instruct base model. It was trained using LoRA (Low-Rank Adaptation) adapters on a comprehensive dataset of 6,112 Sefaria texts, encompassing Torah, Talmud, Mishnah, and various commentaries. The training involved 1,000 steps, resulting in a model specifically optimized for Jewish textual analysis.
Key Capabilities
- Specialized Knowledge: Deep understanding of Torah, Tanakh, Talmudic discourse, Jewish law (Halakha), and philosophy.
- Bilingual Support: Capable of generating accurate and well-sourced responses in both Hebrew and English.
- Textual Analysis: Excels in analyzing Rishonim and Acharonim commentaries and performing Hebrew-English translation within a Jewish context.
- Efficient Inference: Quantized to 4-bit (NF4) for efficient deployment.
Use Cases and Limitations
This model is ideal for educational and research purposes related to Jewish studies, providing structured and knowledgeable responses. However, it is crucial to note its limitations:
- Hallucination Risk: The model may occasionally fabricate sources, necessitating verification of citations.
- Data Scope: Its knowledge is limited to the Sefaria corpus it was trained on.
- Ethical Considerations: It is explicitly not suitable for complex legal decisions (p'sak halacha); users should always consult qualified Rabbis for halachic guidance. For robust, grounded responses, the developers strongly recommend using it with their companion RAG pipeline to prevent fabricated sources and enable conditional refusal.