diff --git a/gptme/eval/main.py b/gptme/eval/main.py index 09de23b5..12b27750 100644 --- a/gptme/eval/main.py +++ b/gptme/eval/main.py @@ -151,14 +151,18 @@ def get_status_emoji(passed, total): return "❌" for model, results in sorted(all_results.items()): - row = [model] + row = [model.replace("openrouter/", "")] for test in headers[1:]: if test in results: passed = results[test]["passed"] total = results[test]["total"] tokens = results[test]["tokens"] status_emoji = get_status_emoji(passed, total) - row.append(f"{status_emoji} {passed}/{total} {tokens}tok") + incl_tokens = True + row.append( + f"{status_emoji} {passed}/{total}" + + (f" {round(tokens / total)}tk" if incl_tokens else "") + ) else: row.append("❓ N/A") table_data.append(row)