# tinker\_cookbook.distillation.train\_off\_policy.Config

## _class_ [**tinker\_cookbook.distillation.train\_off\_policy.Config**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L88)()

Configuration for off-policy distillation with soft teacher targets.

**Fields:**

- [**learning\_rate**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L100) ( _float_)
- [**dataset\_configs**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L101) ( _list[ [DatasetWithTeacher](https://tinker-docs.thinkingmachines.ai/cookbook/api-reference/distillation/datasetwithteacher/) ]_)
- [**model\_name**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L102) ( _str_)
- [**recipe\_name**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L103) ( _str_)
- [**renderer\_name**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L104) ( _str | None_, default: `None`)
- [**lora\_rank**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L105) ( _int_, default: `32`)
- [**n\_teacher\_targets**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L108) ( _int_, default: `20`) – Number of highest-probability teacher tokens per position used as soft targets.
- [**teacher\_concurrency**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L110) ( _int_, default: `64`) – Max concurrent teacher forward passes per batch.
- [**batch\_size**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L114) ( _int_, default: `64`) – Number of examples per training step.
- [**save\_every**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L118) ( _int_, default: `10`) – Checkpointing and logging
- [**eval\_every**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L119) ( _int_, default: `20`)
- [**max\_steps**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L120) ( _int | None_, default: `None`)
- [**load\_checkpoint\_path**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L121) ( _str | None_, default: `None`)
- [**log\_path**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L122) ( _str_)
- [**wandb\_project**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L123) ( _str | None_, default: `None`)
- [**wandb\_name**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L124) ( _str | None_, default: `None`)
- [**base\_url**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L125) ( _str | None_, default: `None`)
- [**ttl\_seconds**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L126) ( _int | None_, default: `604800`) – Server-side checkpoint retention (seconds). None = keep indefinitely.
- [**enable\_trace**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L129) ( _bool_, default: `False`)
- [**evaluator\_builders**](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/distillation/train_off_policy.py#L131) ( _list[SamplingClientEvaluatorBuilder]_, default: `[]`)