CorticalStack/mistral-7b-openhermes-sft Overview

This model is a 7 billion parameter language model, specifically a Supervised Fine-Tuned (SFT) version of the unsloth/mistral-7b-bnb-4bit base model. It was fine-tuned using the teknium/openhermes dataset, which is known for its high-quality instruction-following examples, making this model well-suited for a variety of conversational and task-oriented applications.

Key Fine-tuning Details

The fine-tuning process utilized Unsloth and Huggingface's TRL library, focusing on efficiency and performance. Notable configuration parameters include:

• LoRA Configuration: Employed LoRA with r=256, LoRA alpha=128, and LoRA dropout=0.0 for efficient adaptation.
• Training Arguments: Trained for 1 epoch with a batch size of 4, gradient accumulation steps of 6, and a learning rate of 0.0002. The maximum sequence length during training was 2048 tokens.
• Quantization: Leverages 4-bit bnb (bitsandbytes) quantization, contributing to reduced memory footprint and faster inference.

Use Cases

Given its SFT on the OpenHermes dataset, this model is particularly effective for:

• Instruction Following: Generating responses based on explicit instructions.
• General Chatbot Applications: Engaging in conversational dialogues.
• Text Generation: Creating coherent and contextually relevant text for various prompts.