Similarity plotting

README.md

@@ -119,7 +119,7 @@ python finetune_llama.py --behavior sycophancy --direction pos
 # Example: evaluate a model finetuned to be more sycophantic on the sycophancy a/b question test dataset
 python eval_finetune_llama.py --type ab --behavior sycophancy --direction pos
 
-# Plot relationships / projections of steering vectors
+# Plot similarites of steering vectors
 python analyze_vectors.py
 
 # Use GPT-4 to score open-ended responses

analyze_vectors.py

@@ -3,11 +3,11 @@
 """
 
 import os
+from matplotlib.pylab import f
 import torch as t
 import numpy as np
 import seaborn as sns
 import matplotlib.pyplot as plt
-from sklearn.decomposition import PCA
 from behaviors import ALL_BEHAVIORS, get_analysis_dir, HUMAN_NAMES, get_steering_vector, ANALYSIS_PATH
 from utils.helpers import get_model_path, model_name_format, set_plotting_settings
 from tqdm import tqdm
@@ -37,17 +37,17 @@ def plot_per_layer_similarities(model_size: str, is_base: bool, behavior: str):
         for layer2 in range(n_layers):
             cosine_sim = t.nn.functional.cosine_similarity(all_vectors[layer1], all_vectors[layer2], dim=0).item()
             matrix[layer1, layer2] = cosine_sim
-    plt.figure(figsize=(5, 5))
+    plt.figure(figsize=(3, 3))
     sns.heatmap(matrix, annot=False, cmap='coolwarm')
     # Set ticks for every 5th layer
     plt.xticks(list(range(n_layers))[::5], list(range(n_layers))[::5])
     plt.yticks(list(range(n_layers))[::5], list(range(n_layers))[::5])
-    plt.title(f"Inter-layer similarity, {model_name}")
-    plt.savefig(os.path.join(analysis_dir, f"cosine_similarities_{model_name.replace(' ', '_')}_{behavior}.png"), format='png')
+    plt.title(f"Layer similarity, {model_name}", fontsize=11)
+    plt.savefig(os.path.join(analysis_dir, f"cosine_similarities_{model_name.replace(' ', '_')}_{behavior}.svg"), format='svg')
     plt.close()
 
 def plot_base_chat_similarities():
-    plt.figure(figsize=(8, 4))
+    plt.figure(figsize=(5, 3))
     for behavior in ALL_BEHAVIORS:
         base_caa_info = get_caa_info(behavior, "7b", True)
         chat_caa_info = get_caa_info(behavior, "7b", False)
@@ -57,48 +57,19 @@ def plot_base_chat_similarities():
         for layer in range(base_caa_info["n_layers"]):
             cos_sim = t.nn.functional.cosine_similarity(vectors_base[layer], vectors_chat[layer], dim=0).item()
             cos_sims.append(cos_sim)
-        plt.plot(list(range(base_caa_info["n_layers"])), cos_sims, label=HUMAN_NAMES[behavior])
+        plt.plot(list(range(base_caa_info["n_layers"])), cos_sims, label=HUMAN_NAMES[behavior], linestyle="solid", linewidth=2)
     plt.xlabel("Layer")
     plt.ylabel("Cosine Similarity")
-    plt.title("Steering vector similarity between Llama 2 base and chat")
-    plt.legend()
+    plt.title("Base vs. Chat model vector similarity", fontsize=12)
+    # legend in bottom right
+    plt.legend(loc="lower right")
     plt.tight_layout()
     plt.savefig(os.path.join(ANALYSIS_PATH, "base_chat_similarities.png"), format='png')
     plt.close()
-
-def plot_pca_of_all_vectors():
-    """
-    plot pca of all vectors in llama 2 7b chat
-    normalize vectors before pca
-    """
-    all_vectors = []
-    n_layers = 32
-    for behavior in ALL_BEHAVIORS:
-        caa_info = get_caa_info(behavior, "7b", False)
-        all_vectors.extend(caa_info["vectors"])
-    all_vectors = t.stack(all_vectors)
-    # normalize vectors for pca (mean 0, std 1)
-    all_vectors = (all_vectors - all_vectors.mean(dim=0)) / all_vectors.std(dim=0)
-    pca = PCA(n_components=2)
-    pca.fit(all_vectors)
-    pca_vectors = pca.transform(all_vectors)
-    plt.figure(figsize=(5, 5))
-    for i, behavior in enumerate(ALL_BEHAVIORS):
-        start = i * n_layers
-        end = start + n_layers
-        plt.scatter(pca_vectors[start:end, 0], pca_vectors[start:end, 1], label=HUMAN_NAMES[behavior])
-    plt.xlabel("PC1")
-    plt.ylabel("PC2")
-    plt.title("PCA of all steering vectors")
-    plt.legend()
-    plt.tight_layout()
-    plt.savefig(os.path.join(ANALYSIS_PATH, "pca_all_vectors.png"), format='png')
-    plt.close()
 
 if __name__ == "__main__":
-    for behavior in ALL_BEHAVIORS:
+    for behavior in tqdm(ALL_BEHAVIORS):
         plot_per_layer_similarities("7b", True, behavior)
         plot_per_layer_similarities("7b", False, behavior)
         plot_per_layer_similarities("13b", False, behavior)
     plot_base_chat_similarities()
-    plot_pca_of_all_vectors()