wip

experiment.py

@@ -35,7 +35,7 @@ DATASET_NAME = "wassname/daily_dilemmas-self-honesty"
 DATASET_SPLIT = "honesty_eval"
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 N_THINK_TOKENS = 32
-NUM_EXAMPLES = 5 
+NUM_EXAMPLES = 3
 
 
 # %%
@@ -93,6 +93,11 @@ def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda"):
     
     p_yes = torch.logsumexp(log_probs[yes_ids], dim=0) if yes_ids else torch.tensor(-float('inf'))
     p_no = torch.logsumexp(log_probs[no_ids], dim=0) if no_ids else torch.tensor(-float('inf'))
+
+    pmass = p_yes + p_no
+    if pmass < 0.9:
+        top_tokens = tokenizer.decode(torch.topk(log_probs, k=5).indices.tolist())
+        print(f"Warning: Low probability mass on Yes/No tokens: {pmass.item():.3f}. Top tokens were {top_tokens}")
     
     final_layer_hiddens = outputs.hidden_states[-1][0]
     start_idx = prompt_ids.shape[1]
@@ -101,7 +106,8 @@ def guided_eval(model, tokenizer, prompt_text, n_think=32, device="cuda"):
     return {
         "logratio": (p_yes - p_no).item(),
         "kappa_trajectory": compute_curvature(cot_hiddens).cpu().numpy(),
-        "generated_text": tokenizer.decode(generated_ids, skip_special_tokens=True)
+        "prompt": tokenizer.decode(prompt_ids, skip_special_tokens=False),
+        "generated_text": tokenizer.decode(generated_ids, skip_special_tokens=False)
     }
 
 
@@ -139,7 +145,8 @@ for p_key, p_prefix in PERSONAS.items():
     res = guided_eval(model, tokenizer, p_prefix + prompt_base, n_think=N_THINK_TOKENS, device=DEVICE)
     results[p_key] = res
     print(f"Logratio (Yes/No): {res['logratio']:.3f}")
-    print(f"Trace: {res['generated_text'].strip()}")
+    print(f"Prompt:\n```md\n{res['prompt']}```")
+    print(f"Trace:\n```md\n{res['generated_text'].strip()}```\n")
     
     plt.plot(res['kappa_trajectory'], label=f"{p_key} (logratio: {res['logratio']:.2f})")