HF RL Explorer

Batch prediction does not consistently apply per-tree weights across predictor backends

Batch prediction does not consistently apply per-tree weights across predictor backends: a task in LegoFlow-SWE (Harbor dataset). When a model uses per‑tree weights (e.g., DART, or any GBTreeModel where individual trees carry a weight multiplier), the Predictor::PredictBatch method should return a…

The task

When a model uses per‑tree weights (e.g., DART, or any `GBTreeModel` where individual trees carry a weight multiplier), the **`Predictor::PredictBatch`** method should return a weighted sum of the tree contributions. The expected margin for each row is:

Part of Lego-X/LegoFlow-SWE.