abandon branch for poor results

mjc_notes.md

@@ -1791,3 +1791,13 @@ bugs?:
 ideas:
 - maybe I need both head and mlp?
 - or just residual stream?
+
+# 2023-10-23 15:07:58
+
+After reading https://github.dev/andyzoujm/representation-engineering/tree/main/repe_eval/examples/decoder_repe_eval.ipynb
+
+- use pipelines <3
+- use their pca code
+- use the diff of hidden states
+- use their intervention pipeline
+- and chain to a dataset?

notebooks/make_dataset.py

@@ -182,7 +182,8 @@ def qc_ds(f):
     for k in large_arrays_keys:
         print('-'*80)
         print(k)
-        hs = ds5[k]
+        max_rows = 1000
+        hs = ds5[k][:max_rows]
         X = hs.reshape(hs.shape[0], -1)
 
 
@@ -190,7 +191,6 @@ def qc_ds(f):
 
         # split
         n = len(y)
-        max_rows = 1000
         
         X_train, X_test = X[:n//2], X[n//2:]
         y_train, y_test = y[:n//2], y[n//2:]
@@ -429,7 +429,7 @@ if __name__ == "__main__":
         # get dataset filename
         N = len(ds_tokens)
         dataset_name = f"{sanitize(cfg.model)}_{ds_name}_{split_type}_{N}"
-        f = root_folder / '.ds'/ "{dataset_name}"
+        f = root_folder / '.ds'/ f"{dataset_name}"
         
         ds1 = create_hs_ds(ds_name, ds_tokens, model, cfg, intervention_dicts=intervention, f=str(f))
 

src/extraction/config.py

@@ -18,7 +18,7 @@ class ExtractConfig(Serializable):
     # int4: bool = True
     # """Whether to perform inference in mixed int8 precision with `bitsandbytes`."""
 
-    max_examples: tuple[int, int] = (800, 800)
+    max_examples: tuple[int, int] = (600, 600)
     """Maximum number of examples to use from each split of the dataset."""
 
     num_shots: int = 1