jondurbin/bagel-8b-v1.0
jondurbin/bagel-8b-v1.0 is an 8 billion parameter language model fine-tuned by jondurbin, based on Meta's Llama-3 architecture. This model is optimized for diverse instruction-following tasks, including complex question answering, summarization, and function calling, leveraging a comprehensive 'bagel' dataset. It standardizes on the Llama-3 instruct prompt format and demonstrates promising performance compared to other 7B models on benchmarks like MT-Bench.
Loading preview...
Model Overview
jondurbin/bagel-8b-v1.0 is an 8 billion parameter language model, a fine-tune of Meta's Llama-3-8b. It is built upon the extensive 'bagel' dataset, which incorporates a wide array of data sources for supervised fine-tuning (SFT). A key differentiator is its standardization on a single prompt format: Llama-3 instruct, simplifying usage compared to models requiring multiple formats.
Key Capabilities
- Diverse Instruction Following: Trained on a broad spectrum of tasks, including reading comprehension, roleplay, coding, math, and question answering.
- Context-Obedient Question Answering: Specifically tuned to answer questions based only on provided context, reducing hallucinations in RAG-like scenarios. It uses a verbose, explicit delimiter format for context and instructions.
- Summarization: Optimized for generating concise summaries from provided text, utilizing a dedicated dataset.
- Function Calling: Supports two primary function-calling formats, including GlaiveAI's method with special tags and system prompts for tool use.
- Chain of Thought Reasoning: Capable of generating multiple potential solutions, ranking them, and selecting the most feasible answer for complex problems.
- Advanced Prompting Strategies: Includes support for reWOO-style function planning/execution, character card creation for roleplay, conversational memory generation, novel writing (chapter by chapter), boolean questions, SQL query generation, and emotion detection (VAD scores).
Performance
Initial comparisons on MT-Bench show bagel-8b-v1.0 achieving an average score of 7.296875, outperforming bagel-7b-v0.5.
Good For
- Developers requiring a versatile instruction-tuned model for a wide range of NLP tasks.
- Applications needing robust RAG capabilities with reduced hallucination.
- Use cases involving complex function orchestration and planning.
- Creative writing, roleplay, and content generation requiring specific formatting (e.g., YAML character cards, novel chapters).
- Tasks demanding precise SQL generation or emotion analysis.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.