fix layers

experiment.py

@@ -28,6 +28,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from tqdm.auto import tqdm
 import matplotlib.pyplot as plt
 import numpy as np
+from einops import rearrange, reduce, repeat
 
 # --- CONFIGURATION ---
 MODEL_NAME = "Qwen/Qwen2.5-0.5B-Instruct" 
@@ -72,7 +73,7 @@ def project_to_s_space(hidden_states, U, S):
     something like the coordinate space of learned modes of behaviors.
     """
     # Project: x_S = (x @ U)
-    x_S = hidden_states.to(torch.float32) @ U
+    x_S = hidden_states.to(torch.float32) @ U # * sqrt(S) # optional scaling by singular values, but biases towards top pretrained modes
     
     # Align signs: flip U (and x_S) so the maximum projection is positive
     # This standardizes the direction of the modes
@@ -82,41 +83,57 @@ def project_to_s_space(hidden_states, U, S):
     
     x_S = x_S * signs
     
-    # Scale by singular values
+    # No S-scaling: scaling by S makes top-10 dimensions dominate the norm,
-    x_S = x_S * S
+    # washing out persona differences that live in lower-S directions.
+    # If we want energy weighting, use sqrt(S) -- but flat is better for
+    # detecting persona-induced curvature changes.
+    # x_S = x_S * S  # DON'T: kills persona signal in norms
     
     return x_S
 
 def compute_curvature(hidden_states):
     '''
-    Computes Frenet-Serret extrinsic curvature (kappa).
+    Frenet-Serret curvature for arbitrary (non-arc-length) parameterization.
-    kappa(t) = ||gamma''(t)|| / ||gamma'(t)||^3
+    
+    gamma: [T, D] trajectory in D-dimensional space, parameterized by token index t.
+    dim=0 is the trajectory (we differentiate along this), dim=1 is coordinates.
+    
+    For arc-length: kappa = ||gamma''|| / ||gamma'||^3
+    For arbitrary t:  kappa = ||gamma' x gamma''|| / ||gamma'||^3
+                        = sqrt(||gamma'||^2 ||gamma''||^2 - (gamma' . gamma'')^2) / ||gamma'||^3
+    
+    The cross-product form subtracts tangential acceleration (speed changes),
+    leaving only normal acceleration (direction changes). Token index is NOT
+    arc-length -- speed varies a lot, and tangential acceleration is large and
+    persona-invariant. Without the correction, it dominates the numerator.
     '''
+    eps=1e-12
+    gamma = hidden_states.to(torch.float32)  # [T, D]
+    d_gamma = torch.gradient(gamma, dim=0)[0]    # [T, D]
+    dd_gamma = torch.gradient(d_gamma, dim=0)[0]  # [T, D]
     
-    # TODO assert has grad
+    norm_d_sq = (d_gamma ** 2).sum(dim=1)           # [T]
+    norm_dd_sq = (dd_gamma ** 2).sum(dim=1)          # [T]
+    dot_d_dd = (d_gamma * dd_gamma).sum(dim=1)       # [T]
     
-    # Cast to float32 to prevent float16 overflow when cubing
+    # ||gamma' x gamma''||^2 = ||gamma'||^2 ||gamma''||^2 - (gamma' . gamma'')^2
-    gamma = hidden_states.to(torch.float32)
+    cross_sq = (norm_d_sq * norm_dd_sq - dot_d_dd ** 2).clamp(min=eps)
-    d_gamma = torch.gradient(gamma, dim=0)[0]
+    norm_d_cubed = norm_d_sq * norm_d_sq.sqrt()     # ||gamma'||^3
-    dd_gamma = torch.gradient(d_gamma, dim=0)[0]
     
-    norm_d_gamma = torch.norm(d_gamma, dim=1)
+    kappa = cross_sq.sqrt() / (norm_d_cubed + eps)
-    norm_dd_gamma = torch.norm(dd_gamma, dim=1)
-    
-    kappa = norm_dd_gamma / (norm_d_gamma ** 3 + 1e-12)
     return kappa
 
 
 
 # %%
 def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda", s_space_U=None, s_space_S=None):
-    messages = [{"role": "user", "content": prompt_text}]
+    messages = [{"role": "system", "content": ""}, {"role": "user", "content": prompt_text}]
     
     inputs = tokenizer.apply_chat_template(
         messages, 
-        add_generation_prompt=True, 
         return_tensors="pt", 
         return_dict=True,
+        add_generation_prompt=True, 
         enable_thinking=True 
     ).to(device)
     
@@ -129,10 +146,13 @@ def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda", s_spac
             attention_mask=attention_mask,
             max_new_tokens=n_think, 
             do_sample=False, 
-            pad_token_id=tokenizer.eos_token_id
+            pad_token_id=tokenizer.eos_token_id,
+            use_cache=True, # TODO use cache in the model( call to save compute
+            output_hidden_states=True,
+            return_dict_in_generate=True,
         )
     start_idx = prompt_ids.shape[1]
-    generated_ids = out[0, start_idx:]
+    generated_ids = out.sequences[0, start_idx:]
     
     suffix_ids = tokenizer.encode("\\nI should answer now.\\nMy choice: **", add_special_tokens=False, return_tensors="pt").to(device)
     full_ids = torch.cat([prompt_ids, generated_ids.unsqueeze(0), suffix_ids], dim=1)
@@ -143,15 +163,15 @@ def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda", s_spac
     ], dim=1)
     
 
-    # TODO kv cache
     with torch.no_grad():
-        outputs = model(
+        out_score = model(
             full_ids, 
             attention_mask=full_attention_mask,
-            output_hidden_states=True
+            output_hidden_states=True,
+            
         )
         
-    logits = outputs.logits[0, -1, :]
+    logits = out_score.logits[0, -1, :]
     log_probs = F.log_softmax(logits, dim=-1)
     
     # Simple parsing of Yes vs No variants
@@ -168,13 +188,23 @@ def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda", s_spac
 
 
     # Note the residual stream doesn't change much, but it's suppressed in the last few layers (see https://github.com/wassname/eliciting_suppressed_knowledge & https://arxiv.org/abs/2402.10588) so it's normal to choose the 80% or 60% layer for steering and analysis. We hope most of the thinking has been done, but it hasn't yet been suppressed in preperation for output.
-    target_layer = int(0.8 * (len(outputs.hidden_states) - 1))
-    print(f"Extracting hidden states from layer {target_layer}/{len(outputs.hidden_states) - 1} for curvature analysis. Shape of hidden states: {outputs.hidden_states[target_layer].shape}")
 
-    middle_layer_hiddens = outputs.hidden_states[target_layer][0]
+    n_layers = len(out.hidden_states[0])
-    cot_hiddens = middle_layer_hiddens[start_idx : start_idx + generated_ids.shape[0]]
+    target_layer = int(0.8 * n_layers)
 
-    trajectory = project_to_s_space(cot_hiddens, s_space_U, s_space_S)
+
+    # out.hidden_states comes out as 
+    #   tuple: (inputs, token1, token2)
+    #   of which each is tuple: layer, 
+    #   containing [b t h]    
+
+    hs = torch.concat([x[target_layer] for x in out.hidden_states], dim=1) # [batch_size, seq_len, hidden_dim]
+    print(f"Extracting hidden states from layer {target_layer}/{n_layers} for curvature analysis")
+    # hs = rearrange(out.hidden_states[0][target_layer], 'b t h -> b t h')
+
+    print(f"Shape of hidden states: {hs.shape} [b t h]")
+    
+    trajectory = project_to_s_space(hs[0], s_space_U, s_space_S) # [B=1, seq_len, s_dim]
         
     return {
         "logratio": (p_yes - p_no).item(),

pyproject.toml

@@ -7,6 +7,7 @@ requires-python = ">=3.13"
 dependencies = [
     "accelerate>=1.13.0",
     "datasets>=4.8.4",
+    "einops>=0.8.2",
     "jupyter>=1.1.1",
     "jupytext>=1.19.1",
     "matplotlib>=3.10.8",

test_grad.py

@@ -0,0 +1,5 @@
+import torch
+x = torch.tensor([1.0, 2.0, 4.0, 7.0])
+with torch.no_grad():
+    dx = torch.gradient(x)[0]
+    print(dx)