Model Overview

This model, developed by jtatman, is a LaserRMT fine-tuned version of Aloobun's Reyna-Mini-1.8B-v0.2. It focuses on enhancing conversational abilities and introduces experimental function-calling features. The fine-tuning process utilized the LaserRMT technique, which refines model weights by focusing on the highest-scored weights based on noise-to-ratio analysis, aiming to improve performance while slightly reducing model size.

Key Capabilities

• Conversational AI: Designed for engaging chat and general text generation.
• Function Calling: Includes experimental function-calling capabilities, with plans for future improvements.
• Efficient Fine-tuning: Leverages the LaserRMT technique to avoid training low-performing weights, contributing to a more efficient and potentially smaller model.

Training Details

The model was trained using Axolotl for training and dataset tokenization. The dataset was formatted in ShareGpt conversational style. Hyperparameters included lora_r: 64, lora_alpha: 16, num_epochs: 3, and learning_rate: 0.00025. Evaluation metrics show varying performance across benchmarks like TruthfulQA, Winogrande, and MMLU, with an MMLU accuracy of 0.2533.