jondurbin/airoboros-7b-gpt4-1.4

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Jun 22, 2023License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Cold

The jondurbin/airoboros-7b-gpt4-1.4 model is a full fine-tune of a 7 billion parameter LLaMa model, developed by jondurbin. It was trained using a completely synthetic dataset generated by GPT-4, focusing on enhancing multi-turn conversations, coding in 10 languages, roleplay, jokes, and riddles. This model is specifically optimized for context-obedient question answering, aiming to reduce hallucinations by strictly adhering to provided context, and excels at complex coding instructions and various word games. It is intended for research use only, subject to the LLaMa and OpenAI data usage licenses.

Loading preview...

Overview

jondurbin/airoboros-7b-gpt4-1.4 is a full fine-tune of a 7 billion parameter LLaMa model, built upon the successful GPT-4 series. Its training data is entirely synthetic, generated by GPT-4 via the airoboros project. This iteration introduces several key enhancements over previous versions, making it a versatile tool for various language tasks.

Key Capabilities

  • Enhanced Multi-Turn Conversations: Improved handling and generation of multi-character, multi-turn dialogues.
  • Multi-Language Coding: Includes coding examples in 10 languages, sourced from the rosettacode.org dataset, with a "PLAINFORMAT" option for code-only output.
  • Context-Obedient Question Answering: Specifically tuned to prioritize provided context, minimizing hallucinations by ignoring pre-trained knowledge when explicit context is given. It uses a verbose, delimited format (BEGININPUT, BEGINCONTEXT, ENDCONTEXT, BEGININSTRUCTION, ENDINSTRUCTION) for precise context adherence.
  • Roleplay and Creative Writing: Features more roleplay examples and can generate creative text in specific styles, such as a pirate captain's resignation letter.
  • Word Games & Trivia: Capable of solving anagrams, generating word lists, answering multiple-choice questions, and telling jokes and riddles.

Good for

  • Developers requiring precise context adherence in question-answering systems to reduce factual errors.
  • Generating code snippets and applications across 10 programming languages, with an option for plain code output.
  • Creating engaging multi-character dialogues for interactive applications or storytelling.
  • Experimenting with creative writing prompts and generating text in specific personas.
  • Research into synthetic data training methods and their impact on model capabilities.

Note: This model and its datasets are licensed for research use only due to the underlying LLaMa license and OpenAI's data usage policies, prohibiting commercial use.

Finetunes
4 models