Arli A I_ Mistral Small 22 B Arli A I R P Max v1.1 6.0bpw h6 exl2

Introduction

Mistral-Small-22B-ArliAI-RPMax-v1.1 is a model variant from the RPMax series, designed for creative writing and role-playing (RP) applications. It aims to generate non-repetitive and diverse content by avoiding repeated characters or situations in its training data. This model is intended for personal use due to its restrictive license.

Architecture

The model is based on the mistralai/Mistral-Small-Instruct-2409 architecture and is part of a series with varying parameters ranging from 2B to 70B. The RPMax series focuses on creativity and variety, ensuring models do not latch onto specific personalities or situations.

Training

• Sequence Length: 8192
• Training Duration: Approximately 4 days using 2x3090Ti GPUs
• Epochs: Single epoch to minimize repetition
• QLORA: 64-rank 128-alpha, allowing ~2% of weights to be trainable
• Learning Rate: 0.00001
• Gradient Accumulation: Low value of 32 for enhanced learning

Guide: Running Locally

1. Setup Environment: Install necessary dependencies and set up a Python environment.
2. Download Model: Obtain the model weights from the Hugging Face repository.
3. Install Hugging Face Transformers: Use pip to install the transformers library.
4. Load Model: Use the transformers library to load the model into your environment.
5. Run Inference: Input prompts using the suggested Mistral Instruct Format.

Suggested Cloud GPUs

• NVIDIA A100
• NVIDIA V100
• Google TPU

License

The model is distributed under the Mistral License (MRL), which restricts usage to personal applications. For more information, visit the license link.