jondurbin/bagel-dpo-7b-v0.5
The jondurbin/bagel-dpo-7b-v0.5 is a 7 billion parameter language model developed by jondurbin, fine-tuned from Mistral-7B-v0.2. This model incorporates a Direct Preference Optimization (DPO) pass using the bagel v0.5 dataset, enhancing its ability to follow instructions and generate nuanced responses. It is particularly optimized for diverse tasks including complex question answering, summarization, function calling, and creative writing, supporting a 4096 token context length.
Loading preview...
Model Overview
The jondurbin/bagel-dpo-7b-v0.5 is a 7 billion parameter language model built upon the Mistral-7B-v0.2 architecture. It has undergone a fine-tuning process that includes a Direct Preference Optimization (DPO) pass, utilizing the comprehensive bagel v0.5 dataset. This DPO step is crucial for refining the model's instruction following and response generation capabilities, aiming for more creative, less clichéd, and context-obedient outputs.
Key Capabilities
- Diverse Instruction Following: Trained on a wide array of datasets covering reasoning, coding, reading comprehension, and multi-turn conversations.
- Context-Obedient Question Answering: Specifically tuned to answer questions strictly from provided context, minimizing hallucinations, ideal for RAG applications.
- Advanced Prompting Strategies: Supports specialized formats for summarization, function calling (including reWOO-style planning), chain-of-thought reasoning, and roleplay character card creation.
- Multi-format Prompting: Utilizes four distinct prompt formats (Vicuna, Llama-2, Alpaca, and a modified ChatML) to enhance generalization across various instruction types.
- Specialized Tasks: Includes fine-tuning for SQL query generation, emotion detection (VAD scores), and conversational memory creation.
Good For
- Developers requiring precise RAG: Its context-obedient QA makes it suitable for applications where responses must strictly adhere to provided documents.
- Complex Automation: Excels in function calling and reWOO-style planning for multi-step task execution.
- Creative Content Generation: Capable of generating novel-style writing, roleplay scenarios, and character cards.
- Instruction-tuned Applications: Ideal for a broad range of instruction-following tasks due to its extensive and diverse training data, including academic, coding, and conversational datasets.