fix

app.py

@@ -20,21 +20,18 @@ def cleanup_memory():
         torch.cuda.empty_cache()
     dbg("Memory cleanup complete.")
 
+# KORREKTUR: Die `try...except`-Blöcke werden entfernt, um bei Fehlern einen harten Crash
+# mit vollständigem Traceback in der Konsole zu erzwingen. Kein "Silent Failing" mehr.
+
 def run_single_analysis_display(*args, progress=gr.Progress(track_tqdm=True)):
-    """Wrapper für ein einzelnes manuelles Experiment mit robuster Fehlerbehandlung."""
-    try:
-        results = run_seismic_analysis(*args, progress_callback=progress)
-        stats, deltas = results.get("stats", {}), results.get("state_deltas", [])
-        df = pd.DataFrame({"Internal Step": range(len(deltas)), "State Change (Delta)": deltas})
-        stats_md = f"### Statistical Signature\n- **Mean Delta:** {stats.get('mean_delta', 0):.4f}\n- **Std Dev Delta:** {stats.get('std_delta', 0):.4f}\n- **Max Delta:** {stats.get('max_delta', 0):.4f}\n"
-        serializable_results = json.dumps(results, indent=2, default=str)
-        return f"{results.get('verdict', 'Error')}\n\n{stats_md}", df, serializable_results
-    except Exception:
-        # Im Fehlerfall, gib für jede Komponente einen leeren, aber typ-korrekten Wert zurück.
-        error_message = f"### ❌ Analysis Failed\n```\n{traceback.format_exc()}\n```"
-        return error_message, pd.DataFrame(), "{}"
-    finally:
-        cleanup_memory()
+    """Wrapper für ein einzelnes manuelles Experiment."""
+    results = run_seismic_analysis(*args, progress_callback=progress)
+    stats, deltas = results.get("stats", {}), results.get("state_deltas", [])
+    df = pd.DataFrame({"Internal Step": range(len(deltas)), "State Change (Delta)": deltas})
+    stats_md = f"### Statistical Signature\n- **Mean Delta:** {stats.get('mean_delta', 0):.4f}\n- **Std Dev Delta:** {stats.get('std_delta', 0):.4f}\n- **Max Delta:** {stats.get('max_delta', 0):.4f}\n"
+    serializable_results = json.dumps(results, indent=2, default=str)
+    cleanup_memory()
+    return f"{results.get('verdict', 'Error')}\n\n{stats_md}", df, serializable_results
 
 PLOT_PARAMS = {
     "x": "Step", "y": "Delta", "color": "Experiment",
@@ -43,22 +40,12 @@ PLOT_PARAMS = {
 }
 
 def run_auto_suite_display(model_id, num_steps, seed, experiment_name, progress=gr.Progress(track_tqdm=True)):
-    """
-    Wrapper für die automatisierte Experiment-Suite mit robuster Fehlerbehandlung.
-    """
-    try:
-        summary_df, plot_df, all_results = run_auto_suite(model_id, int(num_steps), int(seed), experiment_name, progress)
-        new_plot = gr.LinePlot(value=plot_df, **PLOT_PARAMS)
-        serializable_results = json.dumps(all_results, indent=2, default=str)
-        return summary_df, new_plot, serializable_results
-    except Exception:
-        # KORREKTUR: Gib für jede Komponente einen typ-korrekten, leeren Wert zurück.
-        # Insbesondere für die JSON-Komponente einen leeren JSON-String.
-        empty_plot = gr.LinePlot(value=pd.DataFrame(), **PLOT_PARAMS)
-        error_string_for_json = json.dumps({"error": traceback.format_exc()}, indent=2)
-        return pd.DataFrame([{"Error": "Experiment failed. See Raw JSON."}]), empty_plot, error_string_for_json
-    finally:
-        cleanup_memory()
+    """Wrapper für die automatisierte Experiment-Suite."""
+    summary_df, plot_df, all_results = run_auto_suite(model_id, int(num_steps), int(seed), experiment_name, progress)
+    new_plot = gr.LinePlot(value=plot_df, **PLOT_PARAMS)
+    serializable_results = json.dumps(all_results, indent=2, default=str)
+    cleanup_memory()
+    return summary_df, new_plot, serializable_results
 
 with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.3") as demo:
     gr.Markdown("# 🧠 Cognitive Seismograph 2.3: Advanced Experiment Suite")

cognitive_mapping_probe/concepts.py

@@ -15,11 +15,22 @@ def _get_last_token_hidden_state(llm: LLM, prompt: str) -> torch.Tensor:
     """Hilfsfunktion, um den Hidden State des letzten Tokens eines Prompts zu erhalten."""
     inputs = llm.tokenizer(prompt, return_tensors="pt").to(llm.model.device)
     with torch.no_grad():
-        # WAHRSCHEINLICHE FEHLERQUELLE: Sicherstellen, dass hier wirklich `llm` steht.
         outputs = llm.model(**inputs, output_hidden_states=True)
     last_hidden_state = outputs.hidden_states[-1][0, -1, :].cpu()
-    assert last_hidden_state.shape == (llm.config.hidden_size,), \
-        f"Hidden state shape mismatch. Expected {(llm.config.hidden_size,)}, got {last_hidden_state.shape}"
+
+    # KORREKTUR: Anstatt auf `llm.config.hidden_size` zuzugreifen, was fragil ist,
+    # leiten wir die erwartete Größe direkt vom Modell selbst ab. Dies ist robust
+    # gegenüber API-Änderungen in `transformers`.
+    expected_size = llm.model.config.hidden_size # Der Name scheint doch korrekt zu sein, aber wir machen es robuster
+    try:
+        # Versuche, die Größe über die Einbettungsschicht zu erhalten, was am stabilsten ist.
+        expected_size = llm.model.get_input_embeddings().weight.shape[1]
+    except AttributeError:
+        # Fallback, falls die Methode nicht existiert
+        expected_size = llm.config.hidden_size
+
+    assert last_hidden_state.shape == (expected_size,), \
+        f"Hidden state shape mismatch. Expected {(expected_size,)}, got {last_hidden_state.shape}"
     return last_hidden_state
 
 @torch.no_grad()
@@ -31,7 +42,7 @@ def get_concept_vector(llm: LLM, concept: str, baseline_words: List[str] = BASEL
     target_hs = _get_last_token_hidden_state(llm, prompt_template.format(concept))
     baseline_hss = []
     for word in tqdm(baseline_words, desc=f"  - Calculating baseline for '{concept}'", leave=False, bar_format="{l_bar}{bar:10}{r_bar}"):
-        baseline_hss.append(_get_last_token_hidden_state(llm, prompt_template.format(word)))
+        baseline_hss.append(_get_last_token_hidden_state(llm, prompt_template.format(concept, word)))
     assert all(hs.shape == target_hs.shape for hs in baseline_hss)
     mean_baseline_hs = torch.stack(baseline_hss).mean(dim=0)
     dbg(f"  - Mean baseline vector computed with norm {torch.norm(mean_baseline_hs).item():.2f}")