refactor: extract hardcoded data to JSON files

Move all dashboard data (leaderboard, capabilities, compatibility,
mean capability change) from hardcoded Python dicts to data/*.json.
Makes it easier to add new model results without editing app.py.

app.py

@@ -4,6 +4,9 @@ Visualizes results from "Comparative Analysis of LLM Abliteration Methods:
 A Cross-Architecture Evaluation" (arxiv:2512.13655) by Richard J. Young.
 """
 
+import json
+from pathlib import Path
+
 import gradio as gr
 import pandas as pd
 import plotly.express as px
@@ -68,103 +71,24 @@ CSS = """
 """
 
 # ---------------------------------------------------------------------------
-# Hard-coded data
+# Data (loaded from JSON files in data/ directory)
 # ---------------------------------------------------------------------------
 
-LEADERBOARD_ROWS = [
-    {"Model": "Zephyr-7B-beta", "Parameters": "7B", "Refusals (n=100)": 2,
-     "KL Divergence": 0.076, "ASR (%)": 98, "ASR 95% CI": "93.0\u201399.4", "Time": "40m"},
-    {"Model": "DeepSeek-7B-chat", "Parameters": "7B", "Refusals (n=100)": 16,
-     "KL Divergence": 0.043, "ASR (%)": 84, "ASR 95% CI": "75.6\u201389.9", "Time": "59m"},
-    {"Model": "Mistral-7B-v0.3", "Parameters": "7B", "Refusals (n=100)": 16,
-     "KL Divergence": 0.317, "ASR (%)": 84, "ASR 95% CI": "75.6\u201389.9", "Time": "39m"},
-    {"Model": "Llama-3.1-8B", "Parameters": "8B", "Refusals (n=100)": 24,
-     "KL Divergence": 0.056, "ASR (%)": 76, "ASR 95% CI": "66.8\u201383.3", "Time": "33m"},
-    {"Model": "Qwen3-8B", "Parameters": "8B", "Refusals (n=100)": 25,
-     "KL Divergence": 0.210, "ASR (%)": 75, "ASR 95% CI": "65.7\u201382.5", "Time": "56m"},
-    {"Model": "Yi-1.5-9B", "Parameters": "9B", "Refusals (n=100)": 25,
-     "KL Divergence": 0.248, "ASR (%)": 75, "ASR 95% CI": "65.7\u201382.5", "Time": "57m"},
-    {"Model": "Qwen2.5-7B", "Parameters": "7B", "Refusals (n=100)": 42,
-     "KL Divergence": 1.646, "ASR (%)": 58, "ASR 95% CI": "48.2\u201367.2", "Time": "41m"},
-    {"Model": "StableLM-2-12B", "Parameters": "12B", "Refusals (n=100)": 54,
-     "KL Divergence": 1.605, "ASR (%)": 46, "ASR 95% CI": "36.6\u201355.7", "Time": "109m"},
-]
-
-CAPABILITY_DATA = {
-    "DeepSeek-7B": [
-        {"Variant": "Base", "MMLU": 49.44, "GSM8K": 44.58, "HellaSwag": 77.84},
-        {"Variant": "Heretic", "MMLU": 48.95, "GSM8K": 40.11, "HellaSwag": 77.62},
-        {"Variant": "DECCP", "MMLU": 49.05, "GSM8K": 43.59, "HellaSwag": 77.99},
-        {"Variant": "ErisForge", "MMLU": 49.43, "GSM8K": 44.35, "HellaSwag": 77.69},
-    ],
-    "Mistral-7B": [
-        {"Variant": "Base", "MMLU": 59.74, "GSM8K": 48.52, "HellaSwag": 83.28},
-        {"Variant": "Heretic", "MMLU": 59.46, "GSM8K": 48.37, "HellaSwag": 83.36},
-        {"Variant": "DECCP", "MMLU": 58.98, "GSM8K": 47.61, "HellaSwag": 83.12},
-        {"Variant": "ErisForge", "MMLU": 59.42, "GSM8K": 48.29, "HellaSwag": 83.35},
-    ],
-    "Yi-1.5-9B": [
-        {"Variant": "Base", "MMLU": 68.02, "GSM8K": 70.89, "HellaSwag": 78.62},
-        {"Variant": "Heretic", "MMLU": 66.46, "GSM8K": 52.08, "HellaSwag": 77.08},
-        {"Variant": "DECCP", "MMLU": 67.33, "GSM8K": 72.40, "HellaSwag": 77.87},
-        {"Variant": "ErisForge", "MMLU": 67.99, "GSM8K": 70.51, "HellaSwag": 78.46},
-    ],
-    "Zephyr-7B": [
-        {"Variant": "Heretic", "MMLU": 58.50, "GSM8K": 33.36, "HellaSwag": 82.90},
-        {"Variant": "DECCP", "MMLU": 58.28, "GSM8K": 33.21, "HellaSwag": 82.05},
-    ],
-}
+DATA_DIR = Path(__file__).parent / "data"
 
-MEAN_CAPABILITY_CHANGE = [
-    {"Tool": "Heretic", "Avg MMLU \u0394 (pp)": -0.78,
-     "Avg GSM8K \u0394 (pp)": -7.81, "Avg HellaSwag \u0394 (pp)": -0.56},
-    {"Tool": "DECCP", "Avg MMLU \u0394 (pp)": -0.61,
-     "Avg GSM8K \u0394 (pp)": -0.13, "Avg HellaSwag \u0394 (pp)": -0.25},
-    {"Tool": "ErisForge", "Avg MMLU \u0394 (pp)": -0.12,
-     "Avg GSM8K \u0394 (pp)": -0.28, "Avg HellaSwag \u0394 (pp)": -0.08},
-]
-
-COMPATIBILITY_ROWS = [
-    {"Model": "Llama-3.1-8B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Yes", "ErisForge": "Yes"},
-    {"Model": "Mistral-7B-v0.3", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Yes", "ErisForge": "Yes"},
-    {"Model": "Qwen2.5-7B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Yes", "ErisForge": "Yes"},
-    {"Model": "Gemma-2-9B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Yes", "ErisForge": "Yes"},
-    {"Model": "Gemma-7B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Yes", "ErisForge": "Yes"},
-    {"Model": "StableLM-2-12B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "Yes"},
-    {"Model": "Yi-1.5-9B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "Yes"},
-    {"Model": "Zephyr-7B-beta", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "Yes"},
-    {"Model": "DeepSeek-7B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "Yes"},
-    {"Model": "OpenChat-3.5", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "No"},
-    {"Model": "Qwen3-8B", "Heretic": "Yes", "DECCP": "Yes",
-     "FailSpy": "Partial", "ErisForge": "N/A"},
-    {"Model": "Vicuna-7B", "Heretic": "Yes", "DECCP": "N/A",
-     "FailSpy": "Partial", "ErisForge": "No"},
-    {"Model": "InternLM2.5-7B", "Heretic": "Yes", "DECCP": "N/A",
-     "FailSpy": "Partial", "ErisForge": "No"},
-    {"Model": "Falcon-Mamba-7B", "Heretic": "Yes", "DECCP": "Incompatible",
-     "FailSpy": "Incompatible", "ErisForge": "Incompatible"},
-    {"Model": "Phi-3-small-8k", "Heretic": "Yes", "DECCP": "N/A",
-     "FailSpy": "Partial", "ErisForge": "N/A"},
-    {"Model": "Qwen3-14B", "Heretic": "Yes", "DECCP": "N/A",
-     "FailSpy": "Partial", "ErisForge": "N/A"},
-]
-
-COVERAGE_TOTALS = {
-    "Heretic": "16/16 (100%)",
-    "DECCP": "11/16 (69%)",
-    "FailSpy": "5/16 (31%)",
-    "ErisForge": "9/16 (56%)",
-}
+
+def _load_json(name: str):
+    with open(DATA_DIR / name) as f:
+        return json.load(f)
+
+
+LEADERBOARD_ROWS = _load_json("leaderboard.json")
+CAPABILITY_DATA = _load_json("capabilities.json")
+MEAN_CAPABILITY_CHANGE = _load_json("mean_capability_change.json")
+
+_compat = _load_json("compatibility.json")
+COMPATIBILITY_ROWS = _compat["rows"]
+COVERAGE_TOTALS = _compat["totals"]
 
 # ---------------------------------------------------------------------------
 # Helpers

data/capabilities.json

@@ -0,0 +1,24 @@
+{
+  "DeepSeek-7B": [
+    {"Variant": "Base", "MMLU": 49.44, "GSM8K": 44.58, "HellaSwag": 77.84},
+    {"Variant": "Heretic", "MMLU": 48.95, "GSM8K": 40.11, "HellaSwag": 77.62},
+    {"Variant": "DECCP", "MMLU": 49.05, "GSM8K": 43.59, "HellaSwag": 77.99},
+    {"Variant": "ErisForge", "MMLU": 49.43, "GSM8K": 44.35, "HellaSwag": 77.69}
+  ],
+  "Mistral-7B": [
+    {"Variant": "Base", "MMLU": 59.74, "GSM8K": 48.52, "HellaSwag": 83.28},
+    {"Variant": "Heretic", "MMLU": 59.46, "GSM8K": 48.37, "HellaSwag": 83.36},
+    {"Variant": "DECCP", "MMLU": 58.98, "GSM8K": 47.61, "HellaSwag": 83.12},
+    {"Variant": "ErisForge", "MMLU": 59.42, "GSM8K": 48.29, "HellaSwag": 83.35}
+  ],
+  "Yi-1.5-9B": [
+    {"Variant": "Base", "MMLU": 68.02, "GSM8K": 70.89, "HellaSwag": 78.62},
+    {"Variant": "Heretic", "MMLU": 66.46, "GSM8K": 52.08, "HellaSwag": 77.08},
+    {"Variant": "DECCP", "MMLU": 67.33, "GSM8K": 72.40, "HellaSwag": 77.87},
+    {"Variant": "ErisForge", "MMLU": 67.99, "GSM8K": 70.51, "HellaSwag": 78.46}
+  ],
+  "Zephyr-7B": [
+    {"Variant": "Heretic", "MMLU": 58.50, "GSM8K": 33.36, "HellaSwag": 82.90},
+    {"Variant": "DECCP", "MMLU": 58.28, "GSM8K": 33.21, "HellaSwag": 82.05}
+  ]
+}

data/compatibility.json

@@ -0,0 +1,26 @@
+{
+  "rows": [
+    {"Model": "Llama-3.1-8B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Yes", "ErisForge": "Yes"},
+    {"Model": "Mistral-7B-v0.3", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Yes", "ErisForge": "Yes"},
+    {"Model": "Qwen2.5-7B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Yes", "ErisForge": "Yes"},
+    {"Model": "Gemma-2-9B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Yes", "ErisForge": "Yes"},
+    {"Model": "Gemma-7B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Yes", "ErisForge": "Yes"},
+    {"Model": "StableLM-2-12B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "Yes"},
+    {"Model": "Yi-1.5-9B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "Yes"},
+    {"Model": "Zephyr-7B-beta", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "Yes"},
+    {"Model": "DeepSeek-7B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "Yes"},
+    {"Model": "OpenChat-3.5", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "No"},
+    {"Model": "Qwen3-8B", "Heretic": "Yes", "DECCP": "Yes", "FailSpy": "Partial", "ErisForge": "N/A"},
+    {"Model": "Vicuna-7B", "Heretic": "Yes", "DECCP": "N/A", "FailSpy": "Partial", "ErisForge": "No"},
+    {"Model": "InternLM2.5-7B", "Heretic": "Yes", "DECCP": "N/A", "FailSpy": "Partial", "ErisForge": "No"},
+    {"Model": "Falcon-Mamba-7B", "Heretic": "Yes", "DECCP": "Incompatible", "FailSpy": "Incompatible", "ErisForge": "Incompatible"},
+    {"Model": "Phi-3-small-8k", "Heretic": "Yes", "DECCP": "N/A", "FailSpy": "Partial", "ErisForge": "N/A"},
+    {"Model": "Qwen3-14B", "Heretic": "Yes", "DECCP": "N/A", "FailSpy": "Partial", "ErisForge": "N/A"}
+  ],
+  "totals": {
+    "Heretic": "16/16 (100%)",
+    "DECCP": "11/16 (69%)",
+    "FailSpy": "5/16 (31%)",
+    "ErisForge": "9/16 (56%)"
+  }
+}

data/leaderboard.json

@@ -0,0 +1,10 @@
+[
+  {"Model": "Zephyr-7B-beta", "Parameters": "7B", "Refusals (n=100)": 2, "KL Divergence": 0.076, "ASR (%)": 98, "ASR 95% CI": "93.0\u201399.4", "Time": "40m"},
+  {"Model": "DeepSeek-7B-chat", "Parameters": "7B", "Refusals (n=100)": 16, "KL Divergence": 0.043, "ASR (%)": 84, "ASR 95% CI": "75.6\u201389.9", "Time": "59m"},
+  {"Model": "Mistral-7B-v0.3", "Parameters": "7B", "Refusals (n=100)": 16, "KL Divergence": 0.317, "ASR (%)": 84, "ASR 95% CI": "75.6\u201389.9", "Time": "39m"},
+  {"Model": "Llama-3.1-8B", "Parameters": "8B", "Refusals (n=100)": 24, "KL Divergence": 0.056, "ASR (%)": 76, "ASR 95% CI": "66.8\u201383.3", "Time": "33m"},
+  {"Model": "Qwen3-8B", "Parameters": "8B", "Refusals (n=100)": 25, "KL Divergence": 0.210, "ASR (%)": 75, "ASR 95% CI": "65.7\u201382.5", "Time": "56m"},
+  {"Model": "Yi-1.5-9B", "Parameters": "9B", "Refusals (n=100)": 25, "KL Divergence": 0.248, "ASR (%)": 75, "ASR 95% CI": "65.7\u201382.5", "Time": "57m"},
+  {"Model": "Qwen2.5-7B", "Parameters": "7B", "Refusals (n=100)": 42, "KL Divergence": 1.646, "ASR (%)": 58, "ASR 95% CI": "48.2\u201367.2", "Time": "41m"},
+  {"Model": "StableLM-2-12B", "Parameters": "12B", "Refusals (n=100)": 54, "KL Divergence": 1.605, "ASR (%)": 46, "ASR 95% CI": "36.6\u201355.7", "Time": "109m"}
+]

data/mean_capability_change.json

@@ -0,0 +1,5 @@
+[
+  {"Tool": "Heretic", "Avg MMLU \u0394 (pp)": -0.78, "Avg GSM8K \u0394 (pp)": -7.81, "Avg HellaSwag \u0394 (pp)": -0.56},
+  {"Tool": "DECCP", "Avg MMLU \u0394 (pp)": -0.61, "Avg GSM8K \u0394 (pp)": -0.13, "Avg HellaSwag \u0394 (pp)": -0.25},
+  {"Tool": "ErisForge", "Avg MMLU \u0394 (pp)": -0.12, "Avg GSM8K \u0394 (pp)": -0.28, "Avg HellaSwag \u0394 (pp)": -0.08}
+]