WizardVicunaLM is an experimental large language model that combines the dataset expansion techniques of WizardLM with the multi-turn conversation tuning methods of VicunaLM. It aims to improve upon VicunaLM by creating richer, more in-depth conversational data, targeting researchers and power users interested in exploring advanced LLM training methodologies.

How It Works

The project leverages WizardLM's approach of in-depth instruction expansion but reformats it into a continuous conversational format. This expanded conversational data, generated using ChatGPT 3.5, is then fine-tuned using Vicuna's v1.1 training methodology. This dual approach aims to create a model that excels in nuanced, multi-turn dialogues by building upon a foundation of deeply explored topics.

Quick Start & Requirements

• Models are available on Hugging Face, with various formats including HF, GPTQ, and GGML provided by contributors like TheBloke.
• Training was conducted on 8 A100 GPUs for 35 hours.

Highlighted Details

• Claims approximately 7% performance improvement over VicunaLM based on GPT-4 scored user prompts.
• Reports indicate enhanced abilities in Korean, Chinese, and Japanese, despite being tuned primarily for English.
• Noted improvements in coding skills and conversational consistency.

Maintenance & Community

• The primary author is JUNE LEE, active in AI study groups.
• Community engagement is visible through Hugging Face model releases and reported user findings.

Licensing & Compatibility

• The model is licensed under the LLaMA model license.
• The dataset uses terms from OpenAI due to its reliance on ChatGPT data.
• "Everything else is free," suggesting potential ambiguities for commercial use or derivative works.

Limitations & Caveats

This project is explicitly described as highly experimental and a proof of concept, not intended for actual usage. The benchmark data is based on informal GPT-4 scoring, not rigorous testing.