allenai/open-instruct-gpt4-alpaca-7b

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Jun 7, 2023Architecture:Transformer0.0K Cold

The allenai/open-instruct-gpt4-alpaca-7b is a 7 billion parameter LLaMa-based model developed by AllenAI, fine-tuned on the GPT-4 Alpaca dataset. This model is specifically designed for instruction-following tasks, leveraging the quality of GPT-4 generated instructions. It serves as a model diff, requiring recovery with an original LLaMa model, and is suitable for research into instruction tuning on open resources.

Loading preview...

Overview

allenai/open-instruct-gpt4-alpaca-7b is a 7 billion parameter LLaMa model fine-tuned by AllenAI using the high-quality GPT-4 Alpaca dataset. This model is a component of the research presented in the paper "How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources." It is distributed as a model difference (diff), meaning users must combine it with an existing LLaMa base model to reconstruct the full instruction-tuned model.

Key Characteristics & Usage

  • Base Model: Built upon the LLaMa architecture.
  • Training Data: Fine-tuned using the GPT-4 Alpaca dataset, which consists of instructions generated by GPT-4.
  • Distribution: Provided as a model diff, requiring a recovery process with an original LLaMa model in Hugging Face format. Instructions for this process are available in the associated GitHub repository.
  • Input Format: Optimized for a specific input format: \n<|user|>\nYour message here!\n<|assistant|>\n. Including a newline after <|assistant|> is crucial for optimal generation quality.

Performance Highlights

Benchmarking results from the associated paper indicate its performance across various tasks:

  • MMLU (0-shot): 42.6
  • MMLU (5-shot): 38.3
  • GSM Direct: 6.5
  • GSM CoT: 10.0
  • AlpacaFarm vs Davinci-003: 57.0
  • Average Score: 28.3

This model is primarily intended for research and development in instruction tuning, offering insights into the capabilities of models fine-tuned on advanced instruction datasets.