Overview

jondurbin/airoboros-m-7b-3.1.2 is an experimental 7 billion parameter instruction-tuned model based on Mistral-7B-v0.1, developed by jondurbin. It leverages synthetic data from the airoboros-3.1 dataset, which includes enhanced MathJSON, log information extraction, anonymization, chat introspection, multi-step instructions, and de-censorship data. This iteration, 3.1.2, refines hyperparameters for improved quality over previous versions.

Key Capabilities

• Advanced Instruction Following: Excels in understanding and executing complex instructions.
• MathJSON Generation: Can produce structured MathJSON solutions for mathematical problems, enabling deterministic calculations.
• Context-Obedient QA: Trained to answer questions strictly based on provided context, minimizing hallucinations.
• Summarization: Capable of summarizing text inputs effectively.
• Complex Coding: Generates code for intricate requirements, supporting various languages and inline criteria.
• Agent/Function Calling: Supports generating JSON or YAML for function calls based on user input, similar to OpenAI's function calling.
• Chain-of-Thought Reasoning: Provides multiple potential solutions, ranks them by logic, and selects the most feasible answer.
• Multi-step Instructions with Acknowledgement: Handles sequential instructions with explicit acknowledgements.

Important Note on Prompt Format

This model utilizes the Llama-2 chat format, a breaking change from previous Vicuna-style formats. Users must adapt their prompting to this new structure, which is included in the tokenizer config for easy apply_chat_template usage.

Good For

• Applications requiring precise and structured instruction following.
• Mathematical problem-solving with verifiable outputs.
• Code generation and agentic workflows.
• Context-aware question answering where hallucination reduction is critical.