Interface ReinforcementTuningHyperParameters

Optional. Adapter size for Reinforcement Tuning.

Optional. Batch size for the tuning job. How many prompts to process at a train step. If not set, the batch size will be determined automatically.

Optional. How often at steps to save checkpoints during training. If not set, one checkpoint per epoch will be set. total_steps = epoch_count * samples_per_prompt / total_prompts_in_dataset

Optional. Number of training epoches for the tuning job.

Optional. How often at steps to evaluate the tuning job during training. If not set, evel will be run per epoch. total_steps = epoch_count * samples_per_prompt / total_prompts_in_dataset

Learning rate multiplier for Reinforcement Learning.

Optional. The maximum number of tokens to generate per prompt. Default to 32768.

Optional. Number of different responses to generate per prompt during tuning.

Optional. The thinking budget for the tuning job to optimize for (Gemini 2.5 only). * -1 means dynamic thinking * 0 means no thinking * > 0 means thinking budget in tokens If not set, default to -1 (dynamic thinking).

Indicates the maximum thinking depth during tuning. Starting from Gemini 3.5 models, the old thinking_budget will no longer be supported and will result in a user error if set. Instead, users should use the thinking_level parameter to control the maximum thinking depth.