Add Pile Uncopyrighted source dataset (#36)

sparse_autoencoder/source_data/pile_uncopyrighted.py

@@ -0,0 +1,112 @@
+"""The Pile Uncopyrighted Dataset."""
+from typing import TypedDict, final
+
+from transformers import PreTrainedTokenizerFast
+
+from sparse_autoencoder.source_data.abstract_dataset import (
+    PreprocessTokenizedPrompts,
+    SourceDataset,
+)
+
+
+class PileUncopyrightedSourceDataBatch(TypedDict):
+    """Pile Uncopyrighted Source Data.
+
+    https://huggingface.co/datasets/monology/pile-uncopyright
+    """
+
+    text: list[str]
+    meta: list[dict[str, dict[str, str]]]
+
+
+@final
+class PileUncopyrightedDataset(SourceDataset[PileUncopyrightedSourceDataBatch]):
+    """The Pile Uncopyrighted Dataset.
+
+    https://huggingface.co/datasets/monology/pile-uncopyrighted
+    """
+
+    tokenizer: PreTrainedTokenizerFast
+
+    def preprocess(
+        self,
+        source_batch: PileUncopyrightedSourceDataBatch,
+        *,
+        context_size: int,
+    ) -> PreprocessTokenizedPrompts:
+        """Preprocess a batch of prompts.
+
+        For each prompt's `text`, tokenize it and chunk into a list of tokenized prompts of length
+        `context_size`. For the last item in the chunk, throw it away if the length is less than
+        `context_size` (i.e. if it would otherwise require padding). Then finally flatten all
+        batches to a single list of tokenized prompts.
+
+        Args:
+            source_batch: A batch of source data. For example, with The Pile dataset this would be a
+                dict including the key "text" with a value of a list of strings (not yet tokenized).
+            context_size: The context size to use when returning a list of tokenized prompts.
+        """
+        prompts: list[str] = source_batch["text"]
+
+        tokenized_prompts = self.tokenizer(prompts)
+
+        # Chunk each tokenized prompt into blocks of context_size, discarding the last block if too
+        # small.
+        context_size_prompts = []
+        for encoding in list(tokenized_prompts["input_ids"]):  # type: ignore
+            chunks = [
+                encoding[i : i + context_size]
+                for i in range(0, len(encoding), context_size)
+                if len(encoding[i : i + context_size]) == context_size
+            ]
+            context_size_prompts.extend(chunks)
+
+        return {"input_ids": context_size_prompts}
+
+    def __init__(
+        self,
+        tokenizer: PreTrainedTokenizerFast,
+        context_size: int = 250,
+        buffer_size: int = 1000,
+        preprocess_batch_size: int = 1000,
+        dataset_path: str = "monology/pile-uncopyrighted",
+        dataset_split: str = "train",
+    ):
+        """Initialize the Pile Uncopyrighted dataset.
+
+        Example:
+            >>> from transformers import PreTrainedTokenizerFast
+            >>> tokenizer = PreTrainedTokenizerFast.from_pretrained("gpt2")
+            >>> data = PileUncopyrightedDataset(
+            ...     tokenizer=tokenizer
+            ... )
+            >>> first_item = next(iter(data))
+            >>> len(first_item["input_ids"])
+            250
+
+        Args:
+            tokenizer: The tokenizer to use to tokenize the prompts.
+            context_size: The context size to use when returning a list of tokenized prompts.
+                *Towards Monosemanticity: Decomposing Language Models With Dictionary Learning* used
+                a context size of 250.
+            buffer_size: The buffer size to use when shuffling the dataset. As the dataset is
+                streamed, this just pre-downloads at least `buffer_size` items and then shuffles
+                just that buffer. Note that the generated activations should also be shuffled before
+                training the sparse autoencoder, so a large buffer may not be strictly necessary
+                here. Note also that this is the number of items in the dataset (e.g. number of
+                prompts) and is typically significantly less than the number of tokenized prompts
+                once the preprocessing function has been applied.
+            preprocess_batch_size: The batch size to use just for preprocessing the dataset (e.g.
+                tokenizing prompts).
+            dataset_path: The path to the dataset on Hugging Face.
+            dataset_split: Dataset split (e.g. `train`).
+        """
+        self.tokenizer = tokenizer
+
+        super().__init__(
+            dataset_path=dataset_path,
+            dataset_split=dataset_split,
+            context_size=context_size,
+            buffer_size=buffer_size,
+            preprocess_batch_size=preprocess_batch_size,
+        )

sparse_autoencoder/source_data/tests/test_pile_uncopyrighted.py

@@ -0,0 +1,19 @@
+"""Pile Uncopyrighted Dataset Tests."""
+import pytest
+from transformers import PreTrainedTokenizerFast
+
+from sparse_autoencoder.source_data.pile_uncopyrighted import PileUncopyrightedDataset
+
+
+@pytest.mark.parametrize("context_size", [50, 250])
+def test_tokenized_prompts_correct_size(context_size: int) -> None:
+    """Test that the tokenized prompts have the correct context size."""
+    tokenizer = PreTrainedTokenizerFast.from_pretrained("gpt2")
+
+    data = PileUncopyrightedDataset(tokenizer=tokenizer, context_size=context_size)
+
+    # Check the first 100 items
+    iterable = iter(data.dataset)
+    for _ in range(100):
+        item = next(iterable)
+        assert len(item["input_ids"]) == context_size