jondurbin/airoboros-7b-gpt4-1.4.1-qlora

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Jun 30, 2023License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Cold

The jondurbin/airoboros-7b-gpt4-1.4.1-qlora is a 7 billion parameter LLaMA-based model fine-tuned using QLoRA with a 4096-token context length. Developed by jondurbin, this model utilizes entirely synthetic training data generated by GPT-4, focusing on enhanced multi-turn conversations, coding examples across 10 languages, and improved roleplay. It excels in context-obedient question answering and diverse creative tasks, including coding, writing, and word games.

Loading preview...

Model Overview

This model, airoboros-7b-gpt4-1.4.1-qlora, is a 7 billion parameter LLaMA-based model developed by jondurbin. It is a QLoRA fine-tune, distinguishing it from full fine-tune versions, and was trained on a completely synthetic dataset generated by GPT-4 (version 1.4.1).

Key Capabilities & Enhancements

  • Synthetic Data Training: Utilizes a unique dataset (jondurbin/airoboros-gpt4-1.4.1) for fine-tuning, created entirely by GPT-4.
  • Improved Conversations: Features fixed and expanded examples of multi-character, multi-turn conversations.
  • Enhanced Coding: Includes coding examples in 10 languages sourced from the rosettacode.org dataset, with a "PLAINFORMAT" option for code-only output.
  • Context-Obedient QA: Specifically tuned to ignore prior knowledge and answer questions strictly based on provided context, reducing hallucinations. It uses a verbose, delimited format for closed-context instructions.
  • Diverse Applications: Demonstrates proficiency in roleplay, jokes, riddles, word games, multiple-choice questions, and creative writing.

Usage & Licensing

This model is intended and licensed for research use only (cc-nc-4.0), due to its LLaMA base and the use of OpenAI-generated data, which restricts commercial application. It is designed to be run with a fork of FastChat, supporting multi-line prompts and optional history management.