Guanaco (33B) API

Guanaco (33B) Description

• Developer/Creator: Tim Dettmers

Overview

The Guanaco-33B is an open-source, high-quality chatbot model developed by finetuning the 33B parameter LLaMA mode using 4-bit QLoRA. It is competitive with commercial chatbots like ChatGPT on benchmarks.

Key Features

• Trained on the multilingual OASST1 dataset, with best performance in high-resource languages
• Finetuned using 4-bit QLoRA with NormalFloat4 datatype for the base model and adapters
• Lightweight adapter-only checkpoints, allowing cheap local experimentation

Intended Use

The Guanaco-33B model is intended for research purposes and may produce problematic outputs. It is available under the Apache 2 license, but requires access to the LLaMA model weights which have additional licensing requirements.

Language Support

The Guanaco-33B model supports multiple languages, with best performance in high-resource languages due to the composition of the OASST1 dataset used for finetuning.

Training Data

The model is finetuned on the OASST1 dataset, a multilingual dataset of open-source assistant conversations. The size and diversity of the dataset allow the model to engage in open-ended conversations on a wide range of topics.

Use cases

• building advanced multilingual chatbots for customer service and educational purposes
• research and experimentation in instruction-tuned large language models
• generating natural and detailed multi-turn conversational dialogues
• role-playing and storytelling applications
• multilingual translation and language understanding tasks
• deployment on mid-to-high end consumer GPUs and research servers for real-time AI applications
• fine-tuning and adapting to domain-specific tasks

Capabilities

• Function calling
• Streaming
• Structured outputs