Hi,
I’m trying to fine-tune a gemini-2.5-flash model using pairs of instruction text & image.
Steps I took:
_ Prepared training & validation JSONL files with 3 examples each, here is the sample of one json json object
{“contents”:[{“parts”:[{“fileData”:{“mimeType”:“image/jpeg”,“fileUri”:“gs://path-to-image”}},{“text”:“”}],“role”:“user”},{“parts”:[{“text”:“<expected structured result”}],“role”:“model”}]}
_ Under Vertex AI > Tuning > click Create Tuned Model > Set the following details:
- tuned model name to “gemini-2.5-flash”
- Region us-central1
- epochs=1(for testing only)
- Set relevant params in Tuning Dataset
_ Start training
Training completed and an endpoint was deployed, but the API field is blank. I notice if I remove the model from the Endpoint object, the actual endpoint url shows up.
Is this expected or a bug?
Can we fine-tune gemini family of models with multi-modal input(or even just text) & deploy to an endpoint?
Has anyone successfully deployed their fine-tuned gemini model into production? I cannot find any documentation showing how to do this, most ends at evaluating the output via Vertex AI UI.
