Model Overview
abideen/gemma-2b-openhermes is a 2.6 billion parameter variant of the Gemma 2B language model, further fine-tuned on the OpenHermes-2.5 preference dataset using QLoRA. This fine-tuning process enhances its ability to follow instructions and engage in conversational interactions, making it suitable for various dialogue-based applications.
Key Capabilities
- Instruction Following: Optimized for understanding and responding to user instructions, leveraging the OpenHermes-2.5 dataset.
- Conversational AI: Designed for generating coherent and contextually appropriate responses in chat-like scenarios.
- Text Generation: Capable of producing English-language text based on diverse prompts.
- Efficient Deployment: As a 2.6B parameter model, it offers a balance between performance and computational efficiency.
Evaluation Highlights
The model's performance was evaluated across several benchmarks, including Nous Benchmark (Agieval, GPT4ALL, BigBench, TruthfulQA) and OpenLLM Benchmark. Notable average scores include:
- Agieval: 24.11
- GPT4ALL: 40.01
- BigBench: 44.75
- TruthfulQA (mc1/mc2): 30.11 / 47.69
- OpenLLM Average: 73.5% (with MMLU at 37.62% and HellaSwag acc_norm at 62.73%)
Training Details
The model was trained with a learning rate of 5e-07, a total batch size of 8 (micro_batch_size 1, gradient_accumulation_steps 8), and 1300 training steps. It utilized the chatml chat template and was configured with QLoRA for efficient fine-tuning.