allenai/open-instruct-stanford-alpaca-13b

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Jun 7, 2023Architecture:Transformer0.0K Cold

The allenai/open-instruct-stanford-alpaca-13b is a 13 billion parameter LLaMa model, fine-tuned by AllenAI on the Stanford Alpaca dataset. This model is specifically designed for instruction-following tasks, leveraging the Alpaca dataset to enhance its ability to respond to user prompts. It serves as a model diff, requiring recovery with an original LLaMa model, and is part of research exploring instruction tuning on open resources.

Loading preview...

Overview

This model, allenai/open-instruct-stanford-alpaca-13b, is a 13 billion parameter LLaMa-based model developed by AllenAI. It has been fine-tuned using the Stanford Alpaca dataset, focusing on instruction-following capabilities. This model is released as a "model diff," meaning it requires an existing LLaMa model in Hugging Face format to be recovered into a fully functional model.

Key Characteristics & Usage

  • Architecture: Based on the LLaMa 13B model, fine-tuned for instruction following.
  • Training: Fine-tuned on the Stanford Alpaca dataset as part of the research detailed in the paper "How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources."
  • Recovery Process: Users must have access to an original LLaMa model and use the provided weight_diff.py script from the allenai/open-instruct codebase to recover the full model.
  • Input Format: Designed to process inputs in a specific user/assistant turn-taking format:
    <|user|>
    Your message here!
    <|assistant|>

Performance Highlights

The model's performance across various benchmarks, as reported in the associated research paper, includes:

  • MMLU: 45.1 (0-shot) and 47.1 (5-shot)
  • GSM: 6.0 (Direct) and 8.0 (CoT)
  • BBH: 35.0 (Direct) and 34.5 (CoT)
  • Codex-Eval: 15.7 (Pass@1) and 27.6 (Pass@10)
  • AlpacaFarm vs Davinci-003: 28.7

This model is suitable for researchers and developers interested in instruction-tuned LLaMa models, particularly those exploring the impact of the Stanford Alpaca dataset on model performance.