Model Overview

abideen/gemma-7b-openhermes is an 8.5 billion parameter language model derived from Google's Gemma 7B. It has undergone further fine-tuning using QLoRA on the OpenHermes-2.5 preference dataset, which is designed to improve conversational quality and alignment. The model utilizes a specific chat template for conversational use, which can be applied via the tokenizer's built-in functionality.

Key Capabilities

• Conversational AI: Optimized for generating coherent and contextually relevant responses in chat-based interactions.
• Text Generation: Capable of producing English-language text from diverse inputs like questions, prompts, or documents.
• Fine-tuned Performance: Leverages the OpenHermes-2.5 dataset to enhance its ability to follow instructions and engage in dialogue.

Evaluation Highlights

Evaluations across various benchmarks indicate its performance:

• Nous Benchmark: Achieved an average of 22.29 on Agieval and 32.00 on GPT4ALL tasks, with a TruthfulQA average of 38.90.
• OpenLLM Benchmark: Demonstrated an average of 73.5% across tasks including arc_challenge, hellaswag, gsm8k, winogrande, and mmlu, with mmlu at 53.62% accuracy.

Training Details

The model was trained with a learning rate of 5e-07, a total batch size of 8, and 1000 training steps, utilizing an Adam optimizer and a cosine learning rate scheduler. It was built with Axolotl, indicating a focus on efficient fine-tuning practices.