I am planning to Finetune the enterprise level dataset on Gemini 2.5 flash. I have following questions related to it.
1.Will the dataset(qna pair) that I will use to finetune be used by standard model for training purposes?
2.I have finetuned the model with specific dataset. Can I finetuned that model with additional dataset(on top of already finetuned model)?
3.I want to use Gemini Live to capture the daily conversation and experience to improve the tacit knowledge capture. Will that data be used by google internally to train its model or will it be safe?
17. Training Restriction. Google will not use Customer Data to train or fine-tune any AI/ML models without Customer’s prior permission or instruction.
2.I have finetuned the model with specific dataset. Can I finetuned that model with additional dataset(on top of already finetuned model)?
Talking about Gemini Supervised Training, once finetuned, it’s done. You can not finetune it again. What you should do instead is to work on your finetuning dataset and evaluate it, don’t forget about checkpoints and prompt design which could help you too.
Also, note that there are some Limitations but it should not be that important for QnA / text use case.
Alternatively, if you want less limitation, you can try custom training but it’s less turnkey and does not concerns Gemini.
3.I want to use Gemini Live to capture the daily conversation and experience to improve the tacit knowledge capture. Will that data be used by google internally to train its model or will it be safe?
You can choose to let Google use your audio and Live recordings to improve its services for everyone in Gemini Apps Activity.
If you turn this setting on:
Google uses your audio recordings and Gemini Live recordings (including audio, video, and screenshares) in Gemini Apps Activity to improve and develop our services (including training generative AI models). This includes recordings already stored in Gemini Apps Activity as well as future ones.
When you finetune Gemini 2.5 Flash, your dataset is not used to train Google’s standard models—it remains private for your finetuned instance. You can perform additional finetuning on top of an already finetuned model with new data. Regarding Gemini Live, data captured for improving tacit knowledge is kept private and not used by Google to train its general models, so it remains secure for your organization.