Model Card: North Mistral 7B - GGML
Model Overview
The North Mistral 7B is part of a series of research experiements into creating Scandinavian LLMs. The current versions are pretrained only, so they will have to be finetuned before used. This repo provides experiemental GGML-versions of these models.
Model Architecture
North Mistral 7B is based on the Mistral architecture, renowned for its effectiveness in capturing complex patterns in large datasets. It utilizes a multi-layer transformer decoder structure.
| version | checkpoint | val_loss |
|---|
| v0.1 | 40k | 1.449 |
Training Data
The model was trained on a diverse dataset primarily in English, Swedish, Danish and Norwegian. A complete datacard will be published later.
Intended Use
This model is intended for developers and researchers only. It is particularly suited for applications requiring understanding and generating human-like text, including conversational agents, content generation tools, and automated translation services.
Limitations
- The model will exhibit biases present in the training data.
- Performance can vary significantly depending on the specificity of the task and the nature of the input data.
- High computational requirements for inference may limit deployment on low-resource devices.
Ethical Considerations
Users are encouraged to evaluate the model carefully in controlled environments before deploying it in critical applications. Ethical use guidelines should be followed to prevent misuse of the model's capabilities, particularly in sensitive contexts.
Licensing
North Mistral 7B is released under the MIT Public License, which allows for both academic and commercial use.
