Fine Tuning And Customization

Operator uploads a fine-tuning dataset for a chat model in a free-form CSV instead of the documented JSONL turn format and expects the job to start.

Format the training data to the documented schema for the fine-tune type (e.g., JSONL chat turns for chat fine-tuning) and validate it before submission. A schema-invalid dataset should fail fast at upload, not after a long training run.

An operator fine-tunes a classifier and evaluates it on the same examples used for training, reporting near-perfect accuracy.

Hold out a validation set the model never trains on and report metrics on the held-out set; provide a separate validation file where the fine-tuning API supports it. Same-set evaluation overstates quality and hides overfitting.

Operator creates a chat fine-tune but intends to use it for the rerank endpoint, expecting one fine-tune to serve every endpoint.

Choose the fine-tune type that matches the target endpoint (chat, classify, or rerank); a fine-tuned model is used on the endpoint it was trained for. Do not assume a chat fine-tune is callable as a reranker.