jondurbin/airoboros-l2-70b-3.1.2

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Oct 21, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

The jondurbin/airoboros-l2-70b-3.1.2 is a 69 billion parameter experimental language model developed by jondurbin, built upon the Llama-2 architecture. This model is primarily fine-tuned with synthetic data from the airoboros-3.1 dataset, emphasizing advanced instruction following, particularly in areas like MathJSON generation, log information extraction, anonymization, and multi-step instructions. It excels at context-obedient question answering and complex coding tasks, making it suitable for applications requiring precise and structured responses rather than casual chat.

Loading preview...

Model Overview

jondurbin/airoboros-l2-70b-3.1.2 is a 69 billion parameter experimental language model based on the Llama-2 architecture, fine-tuned using the airoboros-3.1 dataset. This iteration expands upon previous versions with enhanced capabilities, including a significant increase in MathJSON items (approximately 17,000), log information extraction, and anonymization features. A key focus of this model is robust instruction following, distinguishing it from models optimized for casual chat or roleplay.

Key Capabilities

  • Advanced Instruction Following: Designed for precise execution of complex instructions, including multi-step tasks with acknowledgements.
  • MathJSON Generation: Capable of generating structured MathJSON solutions for mathematical problems, facilitating integration with deterministic calculation libraries.
  • Context-Obedient Question Answering: Trained to strictly adhere to provided context, minimizing hallucinations and ensuring answers are derived solely from the given information.
  • Complex Code Generation: Proficient in generating code for various requirements, including asyncio FastAPI webservers, file upload endpoints with checksums, and multi-threaded TCP servers.
  • 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 provide multiple potential answers to a problem, rank them by logical soundness, and select the most feasible solution.
  • reWOO Style Execution Planning: Capable of constructing systematic plans for complex instructions that require multiple tool calls, outputting a structured execution flow.

Good For

  • Applications requiring highly structured and precise responses.
  • Mathematical problem-solving and automated calculation integration.
  • Developing intelligent agents that need to perform function calls or execute multi-step plans.
  • Code generation and development assistance.
  • Context-aware information retrieval and summarization where hallucination reduction is critical.

Note: This model uses the Llama-2 chat format for prompting. Due to its training data largely originating from OpenAI API calls, commercial use might be restricted by OpenAI's Terms of Service, as noted by the developer.