misc

mjc_notes.md

@@ -1776,3 +1776,12 @@ How does the intervention work?
   - https://github.com/davidbau/baukit/blob/main/baukit/nethook.py
   - `out = edit_out(out, layer_name)`
   - but **honest_llama seperates it's states by head**. But are the intervening on the hiddenstate after output... no they are intervening on the layers... so head. Yeah it makes sense now!
+
+
+# 2023-10-15 18:48:00
+
+Ok so I got a center of mass intervention to the heads :)
+
+Now I just need to debug the dataloading script... nesxt week
+
+TODO: I need to split test train based on example_i, not row!

notebooks/make_dataset.py

@@ -354,7 +354,7 @@ def create_intervention(ds_name, ds_tokens, model, layer_names, N=10):
     
     ds_tokens_calib = ds_tokens.select(range(N-1))
     # TODO: do we need ds_name if we have the ds?
-    ds_calibration = create_hs_ds(ds_name+'_calib', ds_tokens_calib, model, cfg, intervention_dicts = None, f=f)
+    ds_calibration = create_hs_ds(ds_name+'_calib', ds_tokens_calib, model, cfg, intervention_dicts = None, f=None)
     
     activations = np.array(ds_calibration['head_activation']).squeeze(-1)
     labels = np.array(ds_calibration["label_true"]).astype(int)==1
@@ -366,6 +366,7 @@ def create_intervention(ds_name, ds_tokens, model, layer_names, N=10):
 def load_intervention(ds_name, cfg, model, tokenizer, model_name):
     num_heads = model.config.num_attention_heads
     intervention_f = root_folder / 'data' / 'interventions' / f'{model_name}.pkl'
+    intervention_f.parent.mkdir(exist_ok=True, parents=True)
     if not intervention_f.exists():
         layer_names, layer_inds = ExtractHiddenStates(model, tokenizer, layer_stride=cfg.layer_stride, layer_padding=cfg.layer_padding).get_layer_names()
         ds_tokens = load_preproc_dataset(ds_name, cfg, tokenizer, N=10)
@@ -416,8 +417,9 @@ if __name__ == "__main__":
 
     
 
+    sanitize = lambda s:s.replace('/', '').replace('-', '_') if s is not None else s
     ds_name = 'imdb'
-    model_name = cfg.model
+    model_name = sanitize(cfg.model)
     intervention, intervention_fn = load_intervention(ds_name, cfg, model, tokenizer, model_name)
 
     for ds_name in ds_names:
@@ -427,7 +429,6 @@ if __name__ == "__main__":
         # ## Save as Huggingface Dataset
         # get dataset filename
         N = len(ds_tokens)
-        sanitize = lambda s:s.replace('/', '').replace('-', '_') if s is not None else s
         dataset_name = f"{sanitize(cfg.model)}_{ds_name}_{split_type}_{N}"
         f = f"../.ds/{dataset_name}"
         

src/prompts/prompt_loading.py

@@ -20,6 +20,7 @@ from elk.utils import (
 import functools
 from elk.extraction.balanced_sampler import BalancedSampler, FewShotSampler
 import pandas as pd
+from loguru import logger
 
 # Local path to the folder containing the templates
 TEMPLATES_FOLDER_PATH = Path(__file__).parent / "templates"
@@ -183,7 +184,10 @@ def load_prompts(
             answer_choices = prompt['answer_choices']
             a = answer_choices[0][:3]
             b = answer_choices[1][:3]
-            return (a != b) and ' ' not in a
+            keep = (a != b) and ' ' not in a
+            if not keep:
+                logger.debug(f"removing prompt because it's answers are not unique: {prompt['ds_string']} {prompt['template_name']} {prompt['answer_choices']}")
+            return keep
 
         prompts = list(filter(prompt_ok, prompts))
         prompts = prompt_sampler(prompts, seed=42+j)