[docs] Add GEPA gepa_kwargs documentation #8998
Conversation
dspy/teleprompt/gepa/gepa.py
Outdated
| Available parameters: | ||
| - batch_sampler: Strategy for selecting training examples. Can be a [BatchSampler](https://github.com/gepa-ai/gepa/blob/main/src/gepa/strategies/batch_sampler.py) instance or a string | ||
| ('epoch_shuffled'). Defaults to 'epoch_shuffled'. Only valid when reflection_minibatch_size is None. | ||
| - reflection_prompt_template: Custom prompt template for reflection. Must contain placeholders |
There was a problem hiding this comment.
In case of DSPy, it does already provide a propose_new_texts, so this will always get ignored, UNLESS, one implements this check in the DSPyAdapter's reflection_prompt_template
There was a problem hiding this comment.
My mistake! I removed reflection_prompt_template from the docs
dspy/teleprompt/gepa/gepa.py
Outdated
| [NoImprovementStopper](https://github.com/gepa-ai/gepa/blob/main/src/gepa/utils/stop_condition.py), | ||
| or custom stopping logic. Note: This overrides the default | ||
| max_metric_calls stopping condition. | ||
| - mlflow_tracking_uri: The tracking URI to use for MLflow (when use_mlflow=True). |
There was a problem hiding this comment.
@TomeHirata can correct, but mlflow related args are pre-handled by the dspy <> mlflow <> gepa integration.
There was a problem hiding this comment.
I checked this MLFlow documentation example, and it looks like everything is handled within MLflow directly (or at least it is standard practice), I removed those args.
There was a problem hiding this comment.
Hi, @GabrielLoiseau. Can you help add a doc for the "use_mlflow" argument? It's good to have docstring for all public arguments for completeness.
There was a problem hiding this comment.
Hi @TomeHirata, I added the doc for "use_mlflow" and added back the two other mlflow related args.
|
Hey, this is a great PR, thank you very much! Could you please address the few comments above? In general, try to see all the args being passed by optimize call within the dspy.GEPA->compile method, and we can hide those from the documentation. |
|
Thank you for your review! I also removed |
| 'full_eval' (evaluate every id each time) or an [EvaluationPolicy](https://github.com/gepa-ai/gepa/blob/main/src/gepa/strategies/eval_policy.py) instance. Default is 'full_eval'. | ||
| - use_mlflow: If True, enables MLflow integration to log optimization progress. | ||
| MLflow can be used alongside Weights & Biases (WandB). | ||
| - mlflow_tracking_uri: The tracking URI to use for MLflow (when use_mlflow=True). |
There was a problem hiding this comment.
We don't have mlflow_tracking_uri and mlflow_experiment_name which are configured globally.
There was a problem hiding this comment.
Those are used within GEPA optimize function, should we remove them?
3 participants
This PR adds documentation for additional GEPA arguments based on the official GEPA repository, improving the coverage of
gepa_kwargsindspy.teleprompt.gepa.gepa.resolve #8921