WizardLM proposed by researchers at Microsoft | | | | Turtles AI

  New #AI #system #WizardLM from #Microsoft #researchers matches proprietary models like #GPT-3 at #language tasks. Using #innovative #training method, it reaches state-of-the-art #performance on #benchmarks for #understanding, #reasoning, and #codegeneration. Largest version passes 57% of test cases on #HumanEval, around 20 points higher than other open-source models. Now available on #GitHub (click here). A new artificial intelligence system called WizardLM is gaining attention in the machine learning community. Developed by researchers at Microsoft, WizardLM is a family of large language models that have been trained to follow complex instructions across a variety of domains. The core innovation behind WizardLM is a method called Evol-Instruct, which automatically generates challenging and diverse instructions to train the models. By exposing the models to these customized instructions during training, WizardLM develops strong capabilities in interpreting and executing commands. Early results suggest WizardLM significantly outperforms other publicly available systems. The researchers benchmarked WizardLM against models like GPT-3 and Codex on tasks testing natural language understanding, reasoning, and code generation. The largest version of WizardLM matched or exceeded the scores of these proprietary models despite being openly available. The WizardLM family currently includes models optimized for general language tasks, coding, and math. WizardCoder, for example, achieved state-of-the-art results on HumanEval, a popular benchmark for evaluating code generation. The researchers report it passed 57% of test cases, around 20 points higher than any existing open-source model. WizardLM is available on GitHub and can be used for free for academic research purposes. The researchers are currently focusing on improving the Evol-Instruct method and plan to release an updated version of WizardLM soon. They welcome feedback from the community on WizardLM's current capabilities and limitations. Highlights: - WizardLM is a family of large language models trained by researchers at Microsoft to follow complex instructions - It uses a new training method called Evol-Instruct to automatically generate challenging instructions - Early benchmarks show WizardLM matches or beats proprietary models like GPT-3 at language and coding tasks - Largest model passes 57% of test cases on HumanEval code benchmark, around 20 points higher than other open-source models - Now available on GitHub for free academic use; researchers welcome feedback The release of WizardLM represents an exciting development in open-source AI capabilities. While large language models have shown impressive performance on many tasks, they often struggle with complex, multi-step instructions. WizardLM demonstrates how targeted training can unlock new levels of competence without requiring massive increases in model scale. I'm curious to see user feedback on where WizardLM excels versus proprietary systems, and where it still falls short. What unique capabilities would you like to see in the next version? How can the community collaborate to push instruction-following models forward?