Model Overview Card for Command R+

Basic Information

• Model Name: Command R+
• Developer/Creator: Cohere
• Release Date: April 8, 2024
• Version: August 2024
• Model Type: Large Language Model (LLM)

Description

Overview

Command R+ is a cutting-edge large language model designed for enterprise applications, focusing on advanced capabilities such as Retrieval-Augmented Generation (RAG) and multi-step tool use. It is built to enhance efficiency and accuracy in real-world business scenarios.

Key Features

• Parameter Count: 104 billion parameters for high performance.
• Context Length: Supports an impressive context length of 128,000 tokens.
• Multilingual Support: Optimized for performance in ten languages: English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Arabic, and Simplified Chinese.
• Grounded Generation: Capable of generating responses based on external document snippets with citations.
• Tool Use Capabilities: Supports single-step tool use for interacting with external APIs and databases.

Intended Use

Command R+ is tailored for enterprise scenarios such as customer support, finance, human resources, and marketing. It is particularly effective in applications requiring complex reasoning and automation.

Language Support

The model supports multiple languages and has been evaluated in ten primary languages while also incorporating training data from thirteen additional languages.

Technical Details

Architecture

Command R+ utilizes an optimized transformer architecture that is auto-regressive. It employs a combination of supervised fine-tuning (SFT) and preference training to align its outputs with human preferences regarding helpfulness and safety.

Training Data

The model was trained on a diverse dataset that includes a wide variety of texts across multiple domains.

Data Source and Size

Cohere has not specified the exact sources or total size of the training data but emphasizes its diversity to enhance the model's understanding and reduce biases.

Knowledge Cutoff

The model's knowledge is current as of August 2024.

Diversity and Bias

Cohere aims to mitigate biases through diverse training datasets. However, continuous evaluation is necessary to ensure fairness and robustness in real-world applications.

Performance Metrics

• Accuracy: Command R+ achieved a score of 75% on the MMLU benchmark.
• Speed: The model processes tokens at an output speed of approximately 66.5 tokens per second.
• Latency: The average latency for generating responses is around 0.30 seconds, which includes the time taken to receive the first token.
• Throughput Improvement: Compared to its predecessor, Command R+, offers roughly 50% higher throughput while maintaining similar hardware requirements.

Comparison to Other Models

The Command R+ model delivers strong performance across key metrics while maintaining a cost-effective pricing structure. In multilingual translation, Command R+ achieves a BLEU score of 35.9, closely rivaling GPT-4 Turbo’s 36.6 and significantly outperforming Mistral-Large at 31.4. For retrieval-augmented generation (RAG) tasks, it scores 71.5% accuracy, surpassing Mistral-Large (60.7%) and coming close to GPT-4 Turbo’s 76.7%. Additionally, in tool usage, Command R+ achieves a success rate of 74.5%, better than Mistral-Large’s 63.1% and nearly matching GPT-4 Turbo’s 73.7%.

Usage

Code Samples

The model is available on the AI/ML API platform as "cohere/command-r-plus" .

API Documentation

Detailed API Documentation is available here.

Ethical Guidelines

Cohere emphasizes data privacy and security in its model development. Ethical considerations include ensuring that the model does not propagate biases present in the training data.

Licensing

Commercial usage rights are available through Cohere’s API services.

Get Command R+ API here.