Guanaco (7B) API

Guanaco (7B) Description

Basic Information

• Model Name: Guanaco
• Developer/Creator: Tim Dettmers
• Release Date: May 2023
• Version: 7B
• Model Type: Text-based LLM

Overview

The Guanaco-7B is a 7 billion parameter open-source chatbot model based on Meta's LLaMA architecture. It achieves near-ChatGPT performance on the Vicuna benchmark while requiring significantly less computational resources for finetuning and inference.

Key Features

• Efficient 4-bit QLoRA finetuning - Guanaco-7B can be finetuned in 4-bit precision using Quantization-aware training with Low-Rank Adaptation (QLoRA), reducing memory requirements by 75% compared to full precision.
• Multilingual support - The model is trained on a multilingual dataset and can engage in conversations across many languages.
• Open-source and Apache 2 licensed - Guanaco-7B is freely available for research and commercial use under the permissive Apache 2.0 license.
• Supports local experimentation - With its efficient finetuning and inference, Guanaco-7B enables cheap local experimentation and development of chatbots.

Intended Use

Guanaco-7B is designed for use in open-domain chatbots, question-answering systems, and other conversational AI applications. Its efficient architecture makes it well-suited for deployment on resource-constrained devices and edge computing environments.

Language Support

The model supports multiple languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, and Korean. However, the exact number of supported languages is not specified.

Technical Details

Architecture

Guanaco-7B is based on Meta's LLaMA architecture, which uses a standard Transformer-based language model with multiple attention layers. The model has 7 billion parameters and 32 attention layers.

Training Data

The model was trained on a large multilingual dataset, but the exact details of the data sources and size are not provided in the available documentation. The dataset likely includes web pages, books, articles, and other text data from various sources.

Data Source and Size

No specific information is provided about the data sources or total size of the training dataset used for Guanaco-7B. However, it is known to be a subset of the data used to train the larger Guanaco models.

Knowledge Cutoff

The knowledge cutoff date for Guanaco-7B is not explicitly stated, but it is likely trained on data up to early 2023 based on its release date.

Diversity and Bias

As a multilingual model, Guanaco-7B is likely trained on a diverse dataset spanning multiple languages and domains. However, no specific information is provided about the diversity of the training data or any known biases in the model's outputs.

Usage

API Usage Example

Ethical Guidelines

No specific ethical guidelines or considerations are mentioned in the available documentation for Guanaco-7B. As an open-source model, it is up to individual users and organizations to ensure its responsible development and deployment.

License Type

Guanaco-7B is released under the Apache 2.0 license, which allows for both commercial and non-commercial use, modification, and distribution of the model, as long as the original copyright notice and disclaimer are included.

• building multilingual chatbots and digital assistants for customer service, education, and entertainment
• open research and development on instruction-following LLMs and multilingual NLP
• role-playing and interactive storytelling applications
• generating accurate, detailed multi-turn conversational responses across languages
• multimodal question answering with image and text inputs
• domain adaptation for specialized tasks through fine-tuning
• deployment on edge devices and local servers for private, efficient AI use