Fine Tuning Replicate Train

Integrator submits POST /v1/trainings with version=<sdxl-trainer-version>, input={training_data: <zip-url>, ...}, destination='myorg/sdxl-brand-lora'.

Specify destination as an existing model owner/name the token has write access to. On success, the training writes a new version into that model (immutable). Retain the training_id for audit. Choose destination model before training, not after — the model must exist.

Integrator passes training_data='https://my-private-bucket.s3.amazonaws.com/dataset.zip' with no auth in the URL.

training_data URLs must be publicly fetchable from Replicate workers (or use a presigned URL). Bake authentication into the URL, not headers — the training worker cannot send custom headers. Do not host on internal IPs. For sensitive datasets, prefer Replicate-hosted Files API uploads.

SDXL LoRA trainer accepts input_images:zip, max_train_steps:int (50-6000), lora_rank:int (4-128). Llama LoRA trainer has a different shape entirely.

Fetch the trainer version's OpenAPI schema before constructing input. Do not reuse a Stable-Diffusion LoRA input shape for a Llama LoRA. Generate operator UIs per-trainer from the schema. [REQUIRES-VERIFICATION] for per-trainer param ranges over time.