Model Overview
jondurbin/airoboros-l2-13b-3.1.1 is an experimental language model developed by jondurbin, primarily fine-tuned using synthetic data from the Airoboros project. This version is a prompt fix release, addressing an annoying space requirement from its predecessor. It is built on the airoboros-3.1 dataset, which extends the 3.0 dataset with several specialized data types.
Key Capabilities
- Instruction Following: Designed to excel at following complex instructions, making it a general-purpose model with a strong focus on task execution.
- MathJSON Generation: Capable of generating MathJSON solutions for mathematical questions, facilitating integration with deterministic calculation libraries. Requires a low temperature for optimal results.
- Context-Obedient Question Answering: Trained to ignore prior knowledge and strictly use provided context for answers, reducing hallucinations. Utilizes a specific
BEGININPUT/BEGINCONTEXT/ENDINPUT/BEGININSTRUCTION/ENDINSTRUCTION format for closed-context queries. - Summarization: Includes training for summarizing text, using a similar contextual input format.
- Code Generation: Can handle complex coding instructions and generate code in various languages, with an option for plain code output.
- 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: Able to provide multiple potential solutions to a problem, rank them by logical soundness, and select the most feasible answer.
- reWOO Style Execution Planning: Can construct systematic plans for complex instructions requiring multiple tool calls, outputting a sequence of function calls and evidence references.
Good For
- Developers needing a model for precise instruction execution.
- Applications requiring structured mathematical output (MathJSON).
- Use cases where strict adherence to provided context is critical to prevent hallucination.
- Automated code generation and function call orchestration.
- Tasks benefiting from multi-step reasoning and execution planning.