Model Overview
jondurbin/airoboros-l2-7b-2.2 is an experimental language model built using synthetic data generated by the airoboros project. It is a "clean" version, meaning it does not contain de-alignment data, focusing instead on general-purpose instruction following. The model's training data emphasizes instruction response pairs over casual chat or roleplay, with significant portions dedicated to coding, orca-style instructions, and general tasks.
Key Capabilities
- Context-Obedient Question Answering: Trained to prioritize provided context and limit responses to it, reducing hallucinations. It uses a specific
BEGININPUT/BEGINCONTEXT/BEGININSTRUCTION format for closed-context prompts. - Coding: Capable of handling complex coding instructions, including multi-threaded servers and web applications, with an optional
PLAINFORMAT for code-only 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: Can provide multiple potential answers to a problem, rank them by mathematical logic, and select the most feasible one.
- reWOO-style Execution Planning: Generates systematic plans for complex instructions requiring multiple tool calls, outputting a sequence of function calls and their parameters.
Prompt Format
The model uses a new prompt format with newlines separating system, USER, and ASSISTANT roles. While a default system prompt ("A chat.") is common, it also supports custom system prompts for stylized responses or roleplay scenarios. Users are strongly advised to add stopping criteria for USER: or specified names in system prompts to prevent the model from speaking on their behalf.