train_on_policy.Config - Tinker Documentation
tinker_cookbook.distillation.train_on_policy.Config
class tinker_cookbook.distillation.train_on_policy.Config()
Fields:
- learning_rate ( float)
- dataset_configs ( list[ DistillationDatasetConfig])
- model_name ( str)
- recipe_name ( str)
- renderer_name ( str | None, default:
None) - max_tokens ( int)
- temperature ( float, default:
1.0) - compute_post_kl ( bool, default:
False) - evaluator_builders ( list[SamplingClientEvaluatorBuilder], default:
[]) - lora_rank ( int, default:
32) - kl_penalty_coef ( float, default:
1.0) - kl_discount_factor ( float, default:
0.0) - loss_fn ( LossFnType, default:
'importance_sampling') – See https://tinker-docs.thinkingmachines.ai/losses - loss_fn_config ( dict[str, Any] | None, default:
None) - num_substeps ( int, default:
1) – Useful for very large batch sizes. - wandb_project ( str | None, default:
None) - wandb_name ( str | None, default:
None) - log_path ( str)
- base_url ( str | None, default:
None) - enable_trace ( bool, default:
False) - span_chart_every ( int, default:
0) - eval_every ( int, default:
20) - save_every ( int, default:
20) - load_checkpoint_path ( str | None, default:
None) - max_steps ( int | None, default:
None) – Maximum number of training steps. If None, train on the full dataset.