Hello,
I’m looking for an official clarification regarding the use of Gemini-generated outputs as synthetic training data.
I’m developing a small specialized language model called Wynter Wisp. It is intended specifically for text and dictation cleanup, including punctuation restoration, grammar correction, capitalization, formatting, transcription cleanup, handling self-corrections, and preserving the original meaning of dictated text.
I’m considering using Gemini as a teacher model to generate approximately 100,000 input/output examples. The generations would be produced programmatically through the Antigravity CLI, and the resulting Gemini responses would be automatically extracted and stored as training examples.
These examples would then be used to fine-tune a much smaller model, likely under one billion parameters.
Wynter Wisp would not be a general-purpose assistant or intended as a replacement for Gemini. Its scope would be narrowly limited to text and dictation cleanup.
Could someone from Google please clarify whether the applicable terms permit:
-
Programmatically generating Gemini responses through Antigravity for this purpose.
-
Automatically extracting and storing those responses as synthetic training data.
-
Using approximately 100,000 generated examples to train or fine-tune Wynter Wisp.
-
Distributing or commercially using the resulting fine-tuned model.
-
Publicly or openly releasing the resulting model.
I specifically want clarification regarding synthetic-data generation, model training/behavioral distillation, and automated extraction—not merely whether individual Gemini outputs can generally be used commercially.
I previously emailed the Gemini Code Assist feedback address on August 30 but have not received a response, so I’m asking here before proceeding with the dataset generation.
If a Google representative can provide an official interpretation or point me to the appropriate team or policy covering this use case, I would greatly appreciate it.
Thank you.