jondurbin/airoboros-c34b-3.1.2
The jondurbin/airoboros-c34b-3.1.2 is a 34 billion parameter experimental language model developed by jondurbin, fine-tuned with airoboros-3.1 synthetic data. This model, built on the Llama-2 architecture, features a 32K context length and excels at instruction following, including MathJSON generation, context-obedient question answering, summarization, and complex coding tasks. It is particularly optimized for structured outputs like agent/function calling in JSON/YAML and chain-of-thought reasoning.
Loading preview...
Model Overview
jondurbin/airoboros-c34b-3.1.2 is an experimental 34 billion parameter model, fine-tuned using the airoboros-3.1 synthetic dataset. This dataset expands upon airoboros-3.0 with enhanced MathJSON capabilities, log information extraction, anonymization, chat introspection, multi-step instructions with acknowledgment, and de-censorship data. The model utilizes the Llama-2 chat format for prompting and supports a 32K context length.
Key Capabilities
- Instruction Following: Strong emphasis on adhering to complex instructions rather than casual chat.
- MathJSON Generation: Generates structured MathJSON solutions for mathematical problems, compatible with libraries like CortexJS Compute Engine.
- Context-Obedient QA: Designed to answer questions strictly based on provided context, minimizing hallucinations, using a specific delimited format.
- Summarization: Capable of summarizing text, trained on a dedicated summarization dataset.
- Code Generation: Handles complex coding instructions and can output plain code without additional explanations.
- Agent/Function Calling: Generates JSON or YAML for function calls based on user input, similar to OpenAI's function calling.
- Chain-of-Thought Reasoning: Provides multiple potential answers, ranks them by mathematical logic, and selects the most feasible solution.
- reWOO-style Execution Planning: Supports generating multi-step plans for complex tasks requiring tool use, outputting a structured plan for external execution.
Usage Considerations
This model is generally purpose-built but prioritizes instruction following. Due to its training data generation via OpenAI API calls, commercial use may be restricted by OpenAI's Terms of Service. Users should be aware of the underlying Llama-2 license and its non-commercial usage restrictions.