A team of AI researchers has developed a new series of open-source large language models (LLMs) called WizardLM-2, signaling a significant breakthrough in artificial intelligence. Consisting of three models, WizardLM-2 8x22B, WizardLM-2 70B, and WizardLM-2 7B, each model is designed to handle different complex tasks, aiming to enhance machine learning capabilities.
The introduction of WizardLM-2 marks an important advancement in AI, resulting from a year of comprehensive research and development. The researchers focused on enhancing the model’s ability to understand complex instructions. The models have shown superior performance in areas like chat, multilingual processing, reasoning, and serving as an agent, equaling some of the best existing proprietary large language models.
The WizardLM-2 8x22B model has been identified as the most advanced open-source LLM for administering intricate tasks. The WizardLM-2 70B has been noted for its proficiency in reasoning, making it an excellent contender for tasks involving deep cognitive processes. The smaller model, WizardLM-2 7B, however smaller, has shown quick response times and impressive performances, rivalling models ten times its size.
The development of WizardsLM-2 used advanced techniques, comprising a fully AI-powered synthetic training system that uses progressive learning. By using the “AI Align AI” (AAA) framework which encourages a supportive and collaborative learning environment among various cutting-edge LLMs, these models are able to enhance each other’s capabilities through simulated interactions and peer learning.
Rigorous evaluations, including human and automated assessments, were conducted on WizardLM-2 to establish its performance against existing models. The results indicated that WizardLM-2 either matched or exceeded the capabilities of top models such as GPT-4.
The launch of WizardLM-2 represents a landmark for the open-source community, introducing advanced tools previously exclusive to proprietary models. The high performance of WizardLM-2’s models in complex AI tasks and the use of progressive learning and AI co-teaching methods with the AAA framework underpin a revolution in training methodologies. This could result in more efficient and effective model training. By making WizardLM-2 openly available, transparency and collaboration are encouraged in the AI community, fostering further innovation and applications across various fields.
The project specification page and more information about WizardLM-2 are presently being finalized by the development team and will be available soon. Despite this new language model being open-source, it appears to outperform GPT-4, a leading model from OpenAI, making this not only a huge achievement for the creators but also a crucial advancement in the AI community.
