Overview
The jondurbin/airoboros-l2-7b-3.0 is an experimental 7B parameter model built on the Llama-2 architecture, fine-tuned by jondurbin using the airoboros-3.0 dataset. This iteration updates the prompt format to Llama-2 chat style and emphasizes strong instruction following over casual conversation. It integrates novel features like MathJSON for precise mathematical operations and includes human-generated multi-turn roleplay data to improve conversational coherence.
Key Capabilities
- Instruction Following: Heavily focused on accurate and robust instruction execution.
- MathJSON Integration: Supports deterministic mathematical calculations via a structured JSON output, designed to be parsed and executed by external tools like CortexJS Compute Engine.
- Context-Obedient QA: Trained to strictly adhere to provided context for question answering, minimizing hallucinations.
- Summarization: Capable of summarizing provided text inputs using a specific closed-context format.
- Code Generation: Generates code based on complex requirements, including multi-threaded servers and web applications.
- Agent/Function Calling: Can generate 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, and selects the most feasible answer for complex problems.
- reWOO-style Execution Planning: Supports generating systematic plans for multi-tool execution based on a set of available functions.
Good For
- Applications requiring precise instruction following and deterministic outputs.
- Tasks involving mathematical problem formulation and external computation.
- Closed-context question answering where hallucination reduction is critical.
- Generating structured outputs for agentic workflows and function calling.
- Complex code generation and summarization tasks.