Mistral AI has recently upgraded its Mistral 7B series of AI models, introducing Mistral-7B-v0.03 and Mistral-7B-Instruct-v0.3 new versions.

Difference between Mistral 7B and Mistral Instruct 7B

Both models have the same level of intelligence. However, the Mistral-7B v0.3-instruct is fine-tuned to follow instructions, allowing it to perform tasks and answer questions naturally. The base model doesn’t have this capability.

What's Upgraded

The new Mistral-7B-v0.3 model offers significant improvements over its predecessor. It has an extended vocabulary and supports the v3 Tokenizer, enhancing language understanding and generation. Additionally, the ability to call external functions opens up many possibilities for integrating the model into various applications.

Changes in Mistral-7B-v0.3-instruct Compared to Mistral-7B-v0.2-instruct:
– Extended vocabulary to 32,768 tokens
– Supports v3 Tokenizer
– Supports function calling

Changes in Mistral-7B-v0.3 Compared to Mistral-7B-v0.2:
– Extended vocabulary to 32,768 tokens

Extended Vocabulary

One of the key improvements in this latest version is the extended vocabulary. The model now supports 32,768 tokens, a significant increase from the previous version. This expanded vocabulary allows Mistral-7B-Instruct-v0.3 to understand and generate a wider range of words and phrases, enabling it to tackle more complex and diverse language tasks.

Support for v3 Tokenizer

Another notable addition is the support for the v3 Tokenizer. Tokenization is a crucial step in natural language processing, where text is broken down into smaller units called tokens. The v3 Tokenizer offers enhanced performance and compatibility, ensuring that the model can process and understand the input text more efficiently.

Function Calling Capability

Arguably the most exciting feature of Mistral-7B-Instruct-v0.3 is its ability to support function calling. This means that the model can now interact with external functions and APIs, greatly expanding its capabilities. By using function calling, developers can integrate Mistral-7B-Instruct-v0.3 into various applications, allowing it to perform tasks beyond simple text generation.

How to Access Mistral-7B-Instruct-v0.3

There are several methods to install and use Mistral models, and below we will discuss some of the most popular ones.

Option 1: ChatLabs

ChatLabs is all-in-one GenAI playground with access to 30+ best AI models at one place. ChatLabs makes it easy to use Mistral-7B-v0.3, Mistral-7B-Instruct-v0.3, and over 30 other AI models.
Here’s how to use ChatLabs:

1. Visit ChatLabs: Go to the ChatLabs website and log in.

2. Choose Your Model: Click the dropdown menu at the top right and pick Mistral 7B model

3. Explore Their Power: Start using the model you selected.

With ChatLabs, you only need one subscription for all Pro models. With ChatLabs Pro account, you get access to Gemini 1.5 Pro, GPT-4 Turbo, Meta AI LLaMA 3, Opus Claude 3, and many more. Plus, you can search the web, create images, explore the prompt library, and build custom AI assistants.

There's also a Split Screen feature that lets you use and compare two models at the same time.

Option 2: Mistral-inference on Hugging Face

If you want access to Mistral-7B v0.3, you can use the official mistral_inference library, which is a convenient option.

Installation from Hugging Face

pip install mistral_inference

Download from Hugging Face

# Download from Hugging Face
from huggingface_hub import snapshot_download
from pathlib import Path
# Define the path to save the model
mistral_models_path = Path.home().joinpath('mistral_models', '7B-Instruct-v0.3')
mistral_models_path.mkdir(parents=True, exist_ok=True)
# Download the model
snapshot_download(repo_id="mistralai/Mistral-7B-Instruct-v0.3",
                  allow_patterns=["params.json", "consolidated.safetensors", "tokenizer.model.v3"],
                  local_dir=mistral_models_path)

Option 3: OLLaMA

OLLaMA is an open-source library that makes it easy to use large language models. It provides a unified interface for models such as GPT-4, Lla, and T5. OLLaMA simplifies the process, making it an excellent choice for developers.

Key Features of OLLaMA:

– Unified Interface: Offers a consistent and easy-to-use interface for different models, reducing the learning curve.
– Model Compatibility: Supports a variety of popular language models, including GPT-4, Llama, and others, giving developers the flexibility they need.
– Simplified Model Loading: Streamlines the process of loading and initializing models, saving time and effort.

Using OLLaMA:

1. Install OLLaMA:

pip install ollama

2. Load the needed model:

from ollama import OLLaMA model = OLLaMA("gpt-3")

3. To generate text:

prompt = "What is Artificial Intelligence?" response = model.generate(prompt) print(response)

Option 4: LM Studio

LM Studio is another excellent platform for working with large language models. It offers a user-friendly interface and features like fine-tuning, prompt engineering, and model evaluation. LM Studio supports models like Mistral, Phi 3, and Gemma, making it versatile for various tasks.

Key Features of LM Studio:
– User-Friendly Interface: Provides an intuitive web-based interface, accessible to users with different technical backgrounds.
– Fine-Tuning: Allows users to fine-tune models on their datasets for specific tasks.
– Prompt Engineering: Helps design effective prompts to improve the model's output quality.
– Model Evaluation: Offers built-in metrics and visualizations to assess model performance.

Using LM Studio:

1. Sign up on the LM Studio website.

2. Download LM Studio for your device.

3. Create a new project and choose a language model.

4. Upload your dataset for fine-tuning or use provided datasets.

5. Configure model settings like the number of epochs, batch size, and learning rate.

6. Train the model and evaluate its performance using the provided tools.

7. Use the trained model for various tasks like text generation and question answering.

While mistral_inference is specifically designed for Mistral models, OLLaMA and LM Studio offer more flexibility in terms of model selection and customization. Developers can choose the best tool based on factors such as ease of use, model compatibility, required features, and performance needs.

Conclusion

Mistral-7B-Instruct-v0.3 is a significant advancement in large language models. With its extended vocabulary, support for the v3 Tokenizer, and function calling capabilities, it offers improved performance and versatility.

When running Mistral-7B-Instruct-v0.3, developers have several options. The mistral_inference library provides an official approach, while ChatLabs, OLLaMA and LM Studio offer more flexible alternatives. By considering ease of use, compatibility, features, and performance, developers can choose the best tool for their projects.

As natural language processing continues to evolve, models like Mistral 7B will play a crucial role in expanding the possibilities of AI. With its advanced capabilities and flexible running options, it is set to become a valuable tool for researchers, developers, and businesses.

Useful Resources

– Mistral AI official documentation
– Mistral-7B-Instruct-v0.3 repository on HuggingFace
– Mistral-7B-v0.3 repository on HuggingFace
– Mistral Inference library on GitHub
– Ollama Mistral 7B

Related Blog Posts
– How to Run Mistral 8x22B Locally on Your Computer
– Can Mistral AI Access the Internet?