fix: pass DOM tree features to model and fix preprocessing bugs

[LahadMbacke]

DataCollator Does Not Pass DOM Features to the Model

torch_call, tf_call et numpy_call in DataCollatorForDOMNodeMask were returning {input_ids, labels}, silently dropping the five DOM tree positional features:

• node_ids
• parent_node_ids
• sibling_node_ids
• depth_ids
• tag_ids

Impact

In DOMLMEmbeddings.forward(), when these keys are missing, the model falls back to padding values for all tree-position embeddings (P0–P4). As a result, the TreePositionEmbeddings never received the actual DOM structure during training. The model was effectively trained as a standard RoBERTa model, without any awareness of DOM hierarchy or structural relationships.

Fix

After calculating inputs and labels, collect each DOM feature of the batch and collate it in the same way as input_ids :

batch = {"input_ids": inputs, "labels": labels}
for key in ["node_ids", "parent_node_ids", "sibling_node_ids", "depth_ids", "tag_ids"]:
    batch[key] = _torch_collate_batch(
        [e[key] for e in examples], self.tokenizer,
        pad_to_multiple_of=self.pad_to_multiple_of
    )
return batch