jondurbin/airoboros-c34b-2.1
jondurbin/airoboros-c34b-2.1 is a 34 billion parameter instruction-tuned Llama-2 model developed by jondurbin, utilizing synthetic data generated by the Airoboros framework. It features experimental support for role-play (RP) style interactions, detailed writing prompts, and next-chapter generation. The model is optimized for context-obedient question answering, complex coding instructions, agent/function calling, chain-of-thought reasoning, and reWOO-style execution planning.
Loading preview...
Model Overview
jondurbin/airoboros-c34b-2.1 is a 34 billion parameter instruction-tuned Llama-2 model, fine-tuned using synthetic data generated by the Airoboros framework. This version introduces several experimental features and improvements, though it currently has a known prompt formatting bug that will be addressed in a future release.
Key Capabilities
- Experimental RP & GTKM Styles: Includes multi-round chats with emotes for role-play (RP) and a "get to know me" (GTKM) style for character-consistent dialog generation.
- Enhanced Writing: Supports longer, more detailed writing prompts and next-chapter generation, with improved quality from updated generation techniques.
- Context-Obedient QA: Trained to ignore prior knowledge and answer questions strictly based on provided context, reducing hallucinations.
- Advanced Instruction Following: Excels at complex coding instructions, agent/function calling (outputting JSON or YAML), chain-of-thought reasoning, and reWOO-style execution planning.
- De-alignment: Includes a small dataset to reduce censorship from the base models.
Prompt Format & Usage
The model uses a flexible chat-based prompt format (A chat. USER: {prompt} ASSISTANT:), with variations in newline/space usage. Users are advised to add stopping criteria on "USER:" for multi-round conversations. Specific formatting is required for closed-context question answering using BEGININPUT/ENDINPUT delimiters to ensure context adherence.
Licensing
Built on Llama-2, the model adheres to Meta's custom license. The fine-tuning data was generated via OpenAI API calls, leading to ambiguity regarding commercial use due to OpenAI's Terms of Service. Users are advised to exercise caution for commercial applications.