From d13b28dd99ab3f98f9ab5bd90281c2385f62868d Mon Sep 17 00:00:00 2001 From: svlandeg Date: Thu, 15 Jan 2026 22:14:21 +0100 Subject: [PATCH] fix relative_diff calculation --- scripts/tok_eval.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/scripts/tok_eval.py b/scripts/tok_eval.py index 9233d71..5e0a79a 100644 --- a/scripts/tok_eval.py +++ b/scripts/tok_eval.py @@ -213,8 +213,8 @@ def print_comparison(baseline_name, baseline_results, ours_results, all_text): ours_data = ours_results[name] # Calculate relative difference (positive means ours is better, negative means worse) - # Using tokens: fewer tokens is better, so we calculate (baseline_tokens - ours_tokens) / baseline_tokens - relative_diff = ((baseline_data['tokens'] - ours_data['tokens']) / baseline_data['tokens']) * 100 + # Using ratios: higher ratio is better, so we calculate (ours_ratio - baseline_ratio) / baseline_ratio + relative_diff = ((ours_data['ratio'] - baseline_data['ratio']) / baseline_data['ratio']) * 100 # Determine which has better compression (higher ratio = better) if baseline_data['ratio'] > ours_data['ratio']: @@ -256,7 +256,7 @@ for baseline_name in ["GPT-2", "GPT-4"]: for name, text in all_text: baseline_data = baseline_results[name] ours_data = ours_results[name] - relative_diff = ((baseline_data['tokens'] - ours_data['tokens']) / baseline_data['tokens']) * 100 + relative_diff = ((ours_data['ratio'] - baseline_data['ratio']) / baseline_data['ratio']) * 100 lines.append(f"| {name} | {baseline_data['bytes']} | {baseline_data['tokens']} | {baseline_data['ratio']:.2f} | {ours_data['tokens']} | {ours_data['ratio']:.2f} | {relative_diff:+.1f}% |") lines.append("") report_markdown = "\n".join(lines)