I've seen distillation used on these types of datasets (where a question has an exact answer), but what about training on data that doesn't necessary have a correct answer? For instance, what if I want to evaluate a set of data (let's say a user profile) and see how likely they are to use a website? In thst case our input is user data (which can vary greatly between users) and our output is a list of websites that are likely to use. Does that mean we just need more data to fine tune with, with more varied cases?
Yeah distillation should work on non exact. In fact that’s mostly where it is used - ie to distill smaller models like in the Distillation video I made. As to whether more data is needed, it’s not so much a factor of whether the questions have exact answers. Data quantity is dependent on the complexity of patterns you are trying to embed.
Howdy! Try this video: Synthetic Data Generation and Fine tuning (OpenAI GPT4o or Llama 3) ua-cam.com/video/iogrvDu5K0k/v-deo.html It covers evals. And then lemme know if you’re looking for more
Hm.... Maybe you can answer one question for my understanding. With Fine Tuning I have to use the Question - Answer Format to get the Data into the model. But what if I don't have the Answers? Or I don't want to train any answers? For example I have a book and I want the Model to be able to answer to questions about the knowledge from the book. And I want the AI to be able to answer any questions. Also questions I havn't thought of. How can I do this?
Option 1: use gpt-4o to generate the answer (like I showed). This only works if gpt4o knows the answer Option 2: prompt gpt 4o with the question AND the book (or sections from it) when generating the answer. This is what I do in the synthetic data generation video. For questions you haven’t thought of , you can generate them by prompting a model. This is also in the synthetic data video. I’ll add links to the description now.
Very clear and concise video, great work!
very clear and thorough. subscribed!
Thanks
I've seen edge device can do amazing thing with Distillation really cool!
Amazing ❤
I would like to know more about finetuning tool calling as well. How that differs then regular finetuning
Take a look for “trelis function calling” and you’ll find vids
nice
I've seen distillation used on these types of datasets (where a question has an exact answer), but what about training on data that doesn't necessary have a correct answer? For instance, what if I want to evaluate a set of data (let's say a user profile) and see how likely they are to use a website? In thst case our input is user data (which can vary greatly between users) and our output is a list of websites that are likely to use.
Does that mean we just need more data to fine tune with, with more varied cases?
Yeah distillation should work on non exact. In fact that’s mostly where it is used - ie to distill smaller models like in the Distillation video I made.
As to whether more data is needed, it’s not so much a factor of whether the questions have exact answers. Data quantity is dependent on the complexity of patterns you are trying to embed.
Would it be possible to cover on the evaluation side of qna generation ?
Howdy!
Try this video: Synthetic Data Generation and Fine tuning (OpenAI GPT4o or Llama 3)
ua-cam.com/video/iogrvDu5K0k/v-deo.html
It covers evals.
And then lemme know if you’re looking for more
Hm.... Maybe you can answer one question for my understanding. With Fine Tuning I have to use the Question - Answer Format to get the Data into the model. But what if I don't have the Answers? Or I don't want to train any answers? For example I have a book and I want the Model to be able to answer to questions about the knowledge from the book. And I want the AI to be able to answer any questions. Also questions I havn't thought of. How can I do this?
Option 1: use gpt-4o to generate the answer (like I showed). This only works if gpt4o knows the answer
Option 2: prompt gpt 4o with the question AND the book (or sections from it) when generating the answer. This is what I do in the synthetic data generation video.
For questions you haven’t thought of , you can generate them by prompting a model. This is also in the synthetic data video.
I’ll add links to the description now.