WizardLM's team says a third-party AI model stole their work
AI developments are mostly open-source, so AI models are trained on similar training patterns.
3 min. read
Published on
Share this article
Read our disclosure page to find out how can you help Windows Report sustain the editorial team Read more
Key notes
- WizardLM, or WizardCoder, is one of the best AI coders on the market right now.
- WizardLM claims that Phind, a competitor, used the WizardCoder model to train itself, without crediting the original work.
- The company behind Phind denies the claims made by WizardLM.
Remember WizardCoder, the AI coder that we recently covered here at Windows Report? We touted as the best AI coder in the world, back then, and for good reason. WizardCoder, which also goes by the name of WizardLM, is capable of correctly writing blocks of code by itself. And as an AI tool, it can greatly help programmers code faster.
And its reputation has grown so much that it seems WizardLM’s training process has been replicated into a third-party AI model, called Phind. But here’s the catch: the Microsoft-funded team behind WizardLM claims that the Phind team stole all their work, without even crediting them.
If you remember, WizardLM is an open-source model that can be used to train your own AI model. However, AI partnerships are the real driving force behind AI development, and the WizardLM team just wants to be credited when their work is used to train other AI models.
The company behind Phind, however, denies that it has used WizardLM to build the model, even though the WizardLM team has evidence to prove that their work was indeed used to create the third-party Phind model.
WizardLM and Phind: A first documented feud between AI models
- Phind uses a data named WizardCoder-style dataset conducted from the WizardCoder Evol-Instruct method to train their V1 Code Llama model.
- Then they continue to use methods from a WizardCoder model to train their V2 model.
- They do not acknowledge the usage and they claim that they use nothing from WizardCoder.
All of these were spotted and apprehended with pictures, and screenshots detailing the apparent conflict. On the other side, the Phind team says that they didn’t use the WizardLM model to train Phind.
Again, we didn’t use their model. Our v1 model (released before WizardCoder) was trained on a WizardCoder-style dataset that we made ourselves and this was the internal nomenclature for the model.
Phind team
Some people suggested that a partnership or a collaboration between the two parties would do more to advance AI development than arguing about it. Truth is, AI development is mostly open-source, all of the time, so models will have similar if not exact training patterns.
However, if your AI model is trained on another model’s training pattern, then it’s only fair to credit the work. After all, this could be the basis of a long and fruitful partnership.
What do you think about this situation?
