jondurbin/airoboros-l2-7b-gpt4-2.0

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Jul 28, 2023License:otherArchitecture:Transformer0.0K Cold

jondurbin/airoboros-l2-7b-gpt4-2.0 is a 7 billion parameter instruction-tuned Llama-2 model developed by jondurbin, fine-tuned exclusively on synthetic instructions generated by the 0614 version of GPT-4. This model is optimized for context-obedient question answering, complex coding instructions, agent/function calling, chain-of-thought reasoning, and reWOO-style execution planning. It is particularly strong in adhering to provided context and generating structured outputs like JSON or YAML for function calls.

Loading preview...

Overview

jondurbin/airoboros-l2-7b-gpt4-2.0 is a 7 billion parameter instruction-tuned Llama-2 model. It is part of the 2.0 series, which means it was fine-tuned exclusively using synthetic instructions generated by the June 2023 version of GPT-4 via the airoboros project. This model is a full fine-tune, not QLoRA, distinguishing it from some other versions in the series.

Key Capabilities

  • Context-Obedient QA: Trained to strictly adhere to provided context, minimizing hallucinations and ignoring prior knowledge. It uses a specific BEGININPUT/BEGINCONTEXT/BEGININSTRUCTION format for closed-context queries.
  • Complex Coding: Capable of generating code based on detailed requirements, including multi-threaded applications and specific data structures. Supports a "PLAINFORMAT" option for code-only output.
  • Agent/Function Calling: Generates function calls and arguments in JSON or YAML format based on user input, similar to OpenAI's function calling.
  • Chain-of-Thought Reasoning: Can provide multiple potential solutions to a problem, rank them by mathematical logic, and select the most feasible answer.
  • reWOO-style Execution Planning: Supports generating systematic plans for complex instructions that require multiple tool calls, outputting a sequence of actions and evidence references.

Good For

  • Applications requiring strict adherence to provided context for question answering.
  • Generating structured code or function calls.
  • Tasks benefiting from multi-step reasoning and planning.
  • Exploring models fine-tuned on a clean, GPT-4 (0614) exclusive dataset.

Finetunes
1 models