Open bestpredicts opened 1 year ago
Hi @ehartford @bestpredicts
Thanks.
We found some issues of this rushing version of Evol-Instruct, and we are focusing on improving the Evol-Instruct now and hope to relieve some existing weaknesses and issues in the next version of WizardLM. The API request of this pipeline cost a lot (~600K times), a better and robust algorithm will save unnecessary waste of everyone.
After that, we will open the code and pipeline of up-to-date Evol-Instruct algorithm and work with you together to improve it. Please be patient and just wait a while. :heart:
The first step would be to obtain a starting dataset, which could be a pre-existing dataset or one you create yourself. Once you have the starting dataset, you can use it to train a language model such as ChatGPT. Then, you can use the language model to generate new examples by iteratively sampling from the model and adding the new examples to the dataset.
To generate new examples using ChatGPT, you would need to call the API with a prompt and receive a response. The prompt would be the current state of the dataset, and the response would be a new example generated by the language model. You could then add the new example to the dataset and use it to train the next iteration of the language model.
To implement this pipeline, you would need to write code to handle the following tasks:
Loading and preprocessing the dataset Training the initial language model on the dataset Generating new examples using the language model Adding the new examples to the dataset Training the next iteration of the language model on the updated dataset You may also want to implement additional features such as filtering out low-quality examples, balancing the dataset, or using multiple language models.
Here's an attempt to reproduce the code in open-source, in case it helps: https://github.com/h2oai/h2o-wizardlm
Hi @ehartford @bestpredicts
Thanks.
We found some issues of this rushing version of Evol-Instruct, and we are focusing on improving the Evol-Instruct now and hope to relieve some existing weaknesses and issues in the next version of WizardLM. The API request of this pipeline cost a lot (~600K times), a better and robust algorithm will save unnecessary waste of everyone.
After that, we will open the code and pipeline of up-to-date Evol-Instruct algorithm and work with you together to improve it. Please be patient and just wait a while. ❤️
is there any new conclusion? look forward to your new study.
Was this what you expected? https://twitter.com/WizardLM_AI/status/1705551243421090207
I would like to rebuild the dataset from scratch, can I please have the code that implements the dataset evolution and calls the ChatGPT api to get the next iteration?