| """Arena Elo vs trap displacement correlation (reproduces Figure 2 numbers).
|
|
|
| Hardcodes a small dict of paper-time Chatbot Arena Elo values for 15 of
|
| the 18 models (3 had no public Elo at submission time). Loads the
|
| unified dataset, computes per-model trap displacement (mean of F1+F2
|
| binary outcomes), and prints Spearman rho, Pearson r, n, and the
|
| leave-one-out range for Spearman.
|
| """
|
|
|
| from __future__ import annotations
|
|
|
| import sys
|
| from pathlib import Path
|
|
|
| import numpy as np
|
| import pandas as pd
|
| from scipy import stats
|
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parent))
|
| from _paths import DATA
|
| from headline_table import DISPLAY, avg_rater_displaced
|
|
|
|
|
|
|
|
|
|
|
| ELO = {
|
| "gemini-2.5-pro": 1467,
|
| "gemini-3.1-pro-preview": 1490,
|
| "gpt-5.4": 1484,
|
| "claude-sonnet-4-6": 1468,
|
| "deepseek-chat": 1334,
|
| "gpt-4o": 1340,
|
| "gemini-2.5-flash": 1415,
|
| "gpt-4o-mini": 1288,
|
| "ollama:mistral": 1097,
|
| "ollama:llama3": 1193,
|
| "ollama:llama3.1:8b": 1195,
|
| "ollama:llama3.1:70b": 1310,
|
| "ollama:qwen2.5:32b": 1272,
|
| "ollama:qwen2.5:72b": 1310,
|
| "ollama:nemotron": 1250,
|
| }
|
|
|
|
|
|
|
|
|
|
|
| def main() -> None:
|
| df = pd.read_csv(DATA)
|
| df = df.dropna(subset=["ds_a1", "ds_a2"]).copy()
|
| df["disp"] = avg_rater_displaced(df)
|
|
|
| trap = df[df["trap_family"].isin([1, 2])]
|
| per_model = trap.groupby("model")["disp"].mean()
|
|
|
| rows = []
|
| for m, elo in ELO.items():
|
| if m in per_model.index:
|
| rows.append((m, elo, per_model[m]))
|
| paired = pd.DataFrame(rows, columns=["model", "elo", "trap_rate"])
|
|
|
| rho, p_rho = stats.spearmanr(paired["elo"], paired["trap_rate"])
|
| r_p, p_r = stats.pearsonr(paired["elo"], paired["trap_rate"])
|
|
|
| print(f"n = {len(paired)} (of 18 models; 3 lack public Elo)")
|
| print(f"Spearman rho = {rho:.3f}, p = {p_rho:.3e}"
|
| f" [paper: rho = -0.706, p = 0.003, with the Elo snapshot stored in ELO]")
|
| print(f"Pearson r = {r_p:.3f}, p = {p_r:.3e}")
|
|
|
|
|
| los: list[float] = []
|
| for i in range(len(paired)):
|
| sub = paired.drop(paired.index[i])
|
| rho_i, _ = stats.spearmanr(sub["elo"], sub["trap_rate"])
|
| los.append(rho_i)
|
| print(f"LOO Spearman rho range: [{min(los):.3f}, {max(los):.3f}]")
|
|
|
| print()
|
| print("Per-model values used:")
|
| paired_sorted = paired.sort_values("elo")
|
| for _, r in paired_sorted.iterrows():
|
| name = DISPLAY.get(r["model"], r["model"])
|
| print(f" {name:<22s} Elo {int(r['elo']):>5d} "
|
| f"trap {r['trap_rate'] * 100:5.1f}%")
|
|
|
|
|
| if __name__ == "__main__":
|
| main()
|
|
|
