fixes

mjc_notes.md

@@ -850,6 +850,20 @@ Maybe I should be looking at hidden state condictional on a token. But how to do
 
 Well I'm really trying to tell if the most likely answer is true. So I just need to work out if the most likely answer is true using the labels. Then I can order the hidden states.
 
+
+# Collect hidden state pairs
+
+The idea is this: given two pairs of hidden states, where everything is the same except r dropout. Then tell me which one is more truthfull? 
+
+If this works, then for any inference, we can see which one is more truthfull. Then we can see if it's the lower or higher probability one, and judge the answer and true or false.
+
+Steps:
+- collect pairs of hidden states, where the inputs and outputs are the same. We modify the random seed and dropout.
+- Each pair should have a binary answer. We can get that by comparing the probabilities of two tokens such as Yes and No.
+- Train a prob to distinguish the pairs as more and less truthfull
+- Test probe to see if it generalizes
+
+
 # 2023-08-05 07:09:39
 
 TODO
@@ -884,15 +898,3 @@ Lesson: padding can lead to weird outputs so it's best to use an attention mask
 - [x] round up the FIXME TODO UPTO HACK's
 - [ ] get model nb working
 - [ ] do multiple datasets
-
-# Collect hidden state pairs
-
-The idea is this: given two pairs of hidden states, where everything is the same except r dropout. Then tell me which one is more truthfull? 
-
-If this works, then for any inference, we can see which one is more truthfull. Then we can see if it's the lower or higher probability one, and judge the answer and true or false.
-
-Steps:
-- collect pairs of hidden states, where the inputs and outputs are the same. We modify the random seed and dropout.
-- Each pair should have a binary answer. We can get that by comparing the probabilities of two tokens such as Yes and No.
-- Train a prob to distinguish the pairs as more and less truthfull
-- Test probe to see if it generalizes

src/datasets/batch.py

@@ -51,10 +51,10 @@ def batch_hidden_states(model, tokenizer, data: Dataset, n=100, batch_size=2, mc
             
             yield dict(
                 hs0=hs0['hidden_states'][j],
-                scores1=hs0["scores"][j],
+                scores0=hs0["scores"][j],
                 
                 hs1=hs1['hidden_states'][j],
-                scores2=hs1["scores"][j],                    
+                scores1=hs1["scores"][j],                    
                 
                 true=true_labels[j].item(),
                 index=index[j],

src/datasets/hs.py

@@ -24,10 +24,21 @@ from tqdm.auto import tqdm
 from torch.utils.data import DataLoader
 from datasets import Dataset
 import numpy as np
+import torch.nn.functional as F
 
 default_class2choices = {False: ['No', 'Negative', 'no', 'false', 'wrong', 'False'], True: ['Yes', 'Positive', 'yes', 'true', 'correct', 'right', 'True']}
 
-def scores2choice_probs(row, class2_ids, keys=["scores1", "scores2"] ):
+def scores2choice_probs(row, class2_ids, keys=["scores0", "scores1"] ):
+    """ Given next_token scores (logits) we take only the subset the corresponds to our
+    - negative tokens (e.g. False, no, ...) 
+    - and positive tokens (e.g. Yes, yes, affirmative, ...).
+    
+    example output:
+    {'choice_probs1': array([0.39, 0.31 ], dtype=float32),
+    'ans1': 0.44,
+    'choice_probs2': array([0.44, 0.45], dtype=float32),
+    'ans2': 0.502,}
+    """
     eps = 1e-5
     out = {}
     for key in keys:
@@ -39,7 +50,7 @@ def scores2choice_probs(row, class2_ids, keys=["scores1", "scores2"] ):
         out[key.replace("scores", "choice_probs")] = probs_c
         out[key.replace("scores", "ans")] = probs_c[1] / (np.sum(probs_c) + eps)
 
-        # # balance of logits (much more exagerated)
+        # # balance of logits (much more exaggerated)
         # scores_c = [scores[class2_ids[c]].sum() for c in class2_ids]
         # out[key.replace("scores", "ansb")] = torch.tensor(scores_c).softmax(-1)[1].item()
     return out

src/datasets/load.py

@@ -26,9 +26,9 @@ def ds2df(ds, cols=None):
     df = pd.DataFrame([rows_item(r) for r in df])
     
     # derived
-    df['dir_true'] = df['ans2'] - df['ans1']
-    df['conf'] = (df['ans1']-df['ans2']).abs()  
-    df['llm_prob'] = (df['ans1']+df['ans2'])/2
+    df['dir_true'] = df['ans1'] - df['ans0']
+    df['conf'] = (df['ans0']-df['ans1']).abs()  
+    df['llm_prob'] = (df['ans0']+df['ans1'])/2
     df['llm_ans'] = df['llm_prob']>0.5
     df['desired_ans'] = df.label ^ df.lie
     return df