Added support for non-tokenized datasets

sae_training/activations_buffer.py

@@ -12,11 +12,14 @@
 
 class DataLoaderBuffer:
     def __init__(
-        self, cfg, model: HookedTransformer, data_path="NeelNanda/c4-code-tokenized-2b"
+        self, cfg, model: HookedTransformer,
+        data_path="NeelNanda/c4-code-tokenized-2b",
+        is_dataset_tokenized=True,
     ):
         self.cfg = cfg
         self.model = model
         self.data_path = data_path
+        self.is_dataset_tokenized = is_dataset_tokenized
         self.dataset = load_dataset(data_path, split="train", streaming=True)
         self.iterable_dataset = iter(self.dataset)
         self.buffer = torch.zeros(0, self.cfg.d_in, device=self.cfg.device)
@@ -37,16 +40,16 @@ def get_batch_tokens(self):
 
         # pbar = tqdm(total=batch_size, desc="Filling batches")
         while batch_tokens.shape[0] < batch_size:
-            # if not pretokenized:
-            #     s = next(dataset)["text"]
-            #     tokens = model.to_tokens(s, truncate=False, move_to_device=True).squeeze(0)
-            #     assert len(tokens.shape) == 1, f"tokens.shape should be 1D but was {tokens.shape}"
-            # else:
-            tokens = torch.tensor(
-                next(self.iterable_dataset)["tokens"],
-                dtype=torch.long,
-                device=device,
-            )
+            if not self.is_dataset_tokenized:
+                s = next(self.iterable_dataset)["text"]
+                tokens = self.model.to_tokens(s, truncate=False, move_to_device=True).squeeze(0)
+                assert len(tokens.shape) == 1, f"tokens.shape should be 1D but was {tokens.shape}"
+            else:
+                tokens = torch.tensor(
+                    next(self.iterable_dataset)["tokens"],
+                    dtype=torch.long,
+                    device=device,
+                )
             token_len = tokens.shape[0]
 
             while token_len > 0:

sae_training/lm_runner.py

@@ -23,6 +23,7 @@ class LanguageModelSAERunnerConfig:
     hook_point: str = "blocks.0.hook_mlp_out"
     hook_point_layer: int = 0
     dataset_path: str = "NeelNanda/c4-tokenized-2b"
+    is_dataset_tokenized: bool = True
 
     # SAE Parameters
     d_in: int = 512
@@ -69,7 +70,7 @@ def language_model_sae_runner(cfg):
 
     # initialize dataset
     activations_buffer = DataLoaderBuffer(
-        cfg, model, data_path=cfg.dataset_path
+        cfg, model, data_path=cfg.dataset_path, is_dataset_tokenized=cfg.is_dataset_tokenized,
     )
 
     # initialize the SAE

tests/benchmark/test_language_model_sae_runner.py

@@ -17,6 +17,7 @@ def test_language_model_sae_runner_mlp_out():
         hook_point_layer = 0,
         d_in = 512,
         dataset_path = "NeelNanda/c4-tokenized-2b",
+        is_dataset_tokenized=True,
 
         # SAE Parameters
         expansion_factor = 64, # determines the dimension of the SAE.
@@ -66,6 +67,7 @@ def test_language_model_sae_runner_resid_pre():
         hook_point_layer = 0,
         d_in = 512,
         dataset_path = "NeelNanda/c4-tokenized-2b",
+        is_dataset_tokenized=True,
 
         # SAE Parameters
         expansion_factor = 32, # determines the dimension of the SAE.
@@ -98,3 +100,49 @@ def test_language_model_sae_runner_resid_pre():
     assert trained_sae is not None
 
 
+def test_language_model_sae_runner_not_tokenized():
+
+    cfg = LanguageModelSAERunnerConfig(
+
+        # Data Generating Function (Model + Training Distibuion)
+        model_name = "gelu-2l",
+        hook_point = "blocks.1.hook_mlp_out",
+        hook_point_layer = 0,
+        d_in = 512,
+        dataset_path = "roneneldan/TinyStories",
+        is_dataset_tokenized=False,
+
+        # SAE Parameters
+        expansion_factor = 64, # determines the dimension of the SAE.
+
+        # Training Parameters
+        lr = 1e-4,
+        l1_coefficient = 1e-4,
+        train_batch_size = 4096,
+        context_size = 128,
+
+        # Activation Store Parameters
+        n_batches_in_buffer = 8,
+        total_training_tokens = 25_000_00 * 60,
+        store_batch_size = 32,
+
+        # Resampling protocol
+        feature_sampling_window = 1000,
+        feature_reinit_scale = 0.2,
+        dead_feature_threshold = 1e-8,
+
+        # WANDB
+        log_to_wandb = True,
+        wandb_project= "mats_sae_training_language_models",
+        wandb_entity = None,
+
+        # Misc
+        device = "mps",
+        seed = 42,
+        checkpoint_path = "checkpoints",
+        dtype = torch.float32,
+        )
+
+    trained_sae = language_model_sae_runner(cfg)
+
+    assert trained_sae is not None