Model Overview

This model, mistral-7b-sft-constitutional-ai, is a fine-tuned variant of the Mistral-7B-v0.1 base model. It has been specifically trained to incorporate principles of Constitutional AI, aiming to generate safer and more aligned responses.

Key Capabilities

• Constitutional AI Alignment: Fine-tuned on datasets like HuggingFaceH4/cai-conversation-harmless, which emphasizes harmless and ethical conversational interactions.
• Conversational Proficiency: Further enhanced using the HuggingFaceH4/ultrachat_200k dataset, improving its general conversational abilities.
• Mistral 7B Architecture: Benefits from the efficient and performant architecture of the Mistral-7B-v0.1 base model.

Training Details

The model was trained with a learning rate of 2e-05 over 1 epoch, utilizing a total batch size of 256 across 8 GPUs. The training achieved a validation loss of 0.9344, indicating effective learning from the alignment-focused datasets.

Intended Use Cases

This model is particularly suitable for applications requiring:

• Safe Chatbots: Developing conversational agents that adhere to ethical guidelines and avoid generating harmful content.
• Content Moderation: Assisting in filtering or generating responses that align with specific safety policies.
• Research in AI Alignment: Exploring the effectiveness of constitutional AI principles in practical language models.