Model Overview
jondurbin/airoboros-m-7b-3.0 is an experimental 7 billion parameter model based on the Mistral-7B architecture, fine-tuned primarily with synthetic data from the airoboros-3.0 dataset. A key update in this version is the adoption of the Llama-2 chat format for prompting.
Key Capabilities & Features
- Instruction Following: Heavily focused on precise instruction adherence, distinguishing it from models optimized for casual chat.
- MathJSON Integration: Designed to formulate mathematical problems into a parseable MathJSON format, enabling deterministic calculations via external tools like CortexJS Compute Engine.
- Context-Obedient QA: Trained to ignore prior knowledge and strictly use provided context for question answering, significantly reducing hallucinations in closed-context scenarios.
- Summarization: Includes training for effective summarization tasks, using a specific input format.
- Code Generation: Capable of generating code based on complex requirements, supporting various languages and inline criteria.
- Agent/Function Calling: Supports generation of JSON or YAML for function/argument calls, similar to OpenAI's function calling.
- Chain-of-Thought Reasoning: Can provide multiple potential solutions, rank them by logical soundness, and select the most feasible answer.
- reWOO Style Execution Planning: Generates systematic plans for complex instructions requiring multiple tool calls, outputting a sequence of function calls and their parameters.
Good For
- Applications requiring strict instruction following and minimal hallucination.
- Tasks involving mathematical problem formulation and external computation.
- Closed-context question answering where adherence to provided information is critical.
- Automated code generation and function calling scenarios.
- Complex problem-solving requiring multi-step reasoning and planning.