Model Overview

The daisd-ai/anydef-orpo is a 7 billion parameter language model, fine-tuned from the mistralai/Mistral-7B-v0.1 base architecture. This model has been specifically optimized using the ORPO (Odds Ratio Preference Optimization) training method.

Key Capabilities

• Definition Extraction: The model is fine-tuned on the daisd-ai/anydef-kilt-tasks dataset, indicating a specialization in tasks involving the extraction and understanding of definitions.
• Knowledge-Intensive Language Processing: Its training on a KILT-based dataset suggests proficiency in tasks that require accessing and processing factual knowledge.
• Mistral-7B Foundation: Benefits from the strong base capabilities of the Mistral-7B model, including efficient inference and good general language understanding.

Training Details

The model was trained with a learning rate of 5e-06 over 3 epochs, utilizing an Adam optimizer and an inverse square root learning rate scheduler with 100 warmup steps. The training involved a total batch size of 64 across 8 GPUs.

Good For

• Applications requiring precise definition extraction from text.
• Research and development in knowledge-intensive NLP tasks.
• Use cases where a specialized, fine-tuned 7B model offers advantages over general-purpose LLMs for specific knowledge-based queries. Further details on intended uses and limitations are available on the daisd-ai GitHub repository.