Overview

CorticalStack/mistral-7b-dolphin-sft is a 7 billion parameter language model derived from unsloth/mistral-7b-bnb-4bit. It has undergone Supervised Fine-Tuning (SFT) using the cognitivecomputations/dolphin dataset, which is known for its diverse and high-quality conversational data. This fine-tuning process aims to enhance the model's ability to follow instructions and generate coherent, contextually relevant responses.

Key Characteristics

• Base Model: unsloth/mistral-7b-bnb-4bit
• Fine-tuning Dataset: cognitivecomputations/dolphin
• Parameter Count: 7 billion
• Context Length: 8192 tokens
• Training Framework: Utilizes Unsloth and Huggingface's TRL library for efficient fine-tuning.

Fine-tuning Configuration Highlights

• LoRA Parameters: r=256, alpha=128, dropout=0.0 for efficient adaptation.
• Training Epochs: 1 epoch with a batch size of 4 and 6 gradient accumulation steps.
• Learning Rate: 0.0002 with a linear scheduler.
• Max Sequence Length: 2048 during training.
• Quantization: Trained with 4-bit bnb (BitsAndBytes) for reduced memory footprint.

Potential Use Cases

This model is well-suited for applications requiring a compact yet capable instruction-following model, such as:

• Chatbots and conversational AI.
• Text generation based on specific prompts.
• Summarization and question-answering tasks where instruction adherence is crucial.