Chen Qian
Chen Qian
When training a deep learning model, it is useful to save the checkpoint per X steps, so users can restore and resume training. For reference, HuggingFace `MLflowCallback` has this functionality:...
# What does this PR do? Context: MLflow has receiving lots of data... so we want to save some budget. In this PR, if we detect that a metric has...
We are reworking DSPy saving in order to make it more robust: - Introduce whole model saving, users can specify `state_only=False` to save the model architecture along with the state....
Similar to `dspy.Assert`/`dspy.Suggest`, but this is constraint tied to modules. It's less powerful, but sufficient for most existing use cases, and the usage is more lightweighted.
`format_turn` method is unnecessarily complex, we should consider refactoring it and only keep the necessary part. In additions, `format_turn` is an ambiguous name, we should find a better name.
Make DatabricksRM compatible with Mosaic agent framework, which has some requirements on the return value, and require some extra setup code.
Setup list: - Bank77 dataset - Model as a single `dspy.ChainOfThoughtWithHint` - Use pydantic to regulate the signature. - MiproV2 optimizer. - Boosted accuracy from 66% to 77%, on the...
Right now `dspy.Retrieve` is not a useful abstraction, and we have too many prebuilt `dspy.Retrieve` implementations. Although it would be nice to have DSPy play as a centralized hub for...
Add saving tutorial for the reworked DSPy saving. This PR depends on #1889
We are replacing all relative imports by absolute imports to get compatible with Google style guide. This is beneficial for automatic API reference generation, which we are adding now.