TILE#
- class sdm.processing.text.TILE(model: str = 'qwen3-0.6b', postprocess: object = None, *, batch_size: int = 64, max_length: int | None = None, device: device | str | None = None, dtype: dtype = torch.bfloat16)#
Bases:
ProcessorThe text cell embedding processor from the “TILE: Bridging LLMs and Table Foundation Models for Tabular Learning with Strings” paper.
Each row is serialized as one text, one sentence per cell. An LLM embeds this text, and token embeddings are mean-pooled for each text cell. Optional postprocessing transforms these embeddings independently per text column, for example to reduce dimensionality.
- Parameters:
model (str) – The LLM, one of
"llama-3.1-8b","llama-3.1-70b","llama-3.2-1b","llama-3.2-3b","qwen3-0.6b","qwen3-4b", or"qwen3-8b".postprocess (object) – Optional postprocessing routine, applied independently for each text column.
batch_size (int) – The batch size of the LLM. Adjusting the batch size can significantly improve processing speed. The optimal value depends on your hardware, model size, precision, and input length.
max_length (int | None) – Maximum number of tokens per row.
dtype (dtype) – The LLM dtype.
Capabilities#
Handled Semantic Types |
|
Requires Fitting |
instance-specific |