jondurbin/spicyboros-70b-2.2
jondurbin/spicyboros-70b-2.2 is an experimental 69 billion parameter language model, primarily built using synthetic data generated by Airoboros. This model is designed for general-purpose instruction response, with a focus on de-alignment to enable less filtered interactions and outputs. It excels in tasks requiring context-obedient question answering, complex coding instructions, agent/function calling, and chain-of-thought reasoning.
Loading preview...
Model Overview
jondurbin/spicyboros-70b-2.2 is a 69 billion parameter experimental language model, primarily fine-tuned on synthetic data generated by the Airoboros project. A key differentiator of this version is the inclusion of "de-alignment" data, enabling the model to produce less filtered and potentially profane or NSFW content when prompted. Users are explicitly warned about this capability and are responsible for its use.
Key Capabilities
- De-aligned Outputs: Designed to provide less censored responses, including profanity and sensitive content, if requested.
- Instruction Following: Optimized for general-purpose instruction response, with a dataset breakdown showing strong representation in categories like
orca,general,coding,roleplay, andtrivia. - Context-Obedient QA: Features a specific prompt format (
BEGININPUT/BEGINCONTEXT/BEGININSTRUCTION) to ensure the model strictly adheres to provided context and reduces hallucinations. - Advanced Coding: Capable of handling complex coding instructions, including inline criteria and an optional
PLAINFORMATfor direct code output. - Agent/Function Calling: Supports generation of JSON or YAML for function/argument selection based on user input, similar to OpenAI's function calling.
- Chain-of-Thought Reasoning: Can generate multiple potential solutions to a problem, rank them by logical soundness, and select a final answer.
- reWOO-style Execution Planning: Facilitates systematic planning for complex instructions requiring multiple tool calls, outputting a structured plan for external execution.
Usage Considerations
This model is built on Llama-2/CodeLlama and uses data generated via OpenAI API calls. Due to potential ambiguities regarding OpenAI's Terms of Service concerning competitive model training, commercial use is advised with caution. The prompt format has been updated to use newlines between system/USER/ASSISTANT roles.