{ "bench_id": "chat-template-toolcall-bench", "bench_version": "1.0", "measured_at_utc": "2026-11-16T16:46:28Z", "measured_at_ist": "2026-11-16 22:16 IST", "elapsed_sec": 32.2, "transformers_version": "5.22.2", "python_version": "3.13.5", "n_repos_listed": 41, "n_repos_measured": 41, "n_load_errors": 0, "n_unique_templates": 33, "n_cases": 10, "C1_dict_args_pass": 40, "C2_string_args_pass": 3, "C2_string_args_crash": 32, "C2_string_args_clear": 5, "C2_string_args_double": 1, "C2_string_args_dropped": 0, "C10_name_escape_pass": 0, "C10_name_escape_raw": 40, "C10_name_escape_error": 1, "n_dict_pass": 40, "of_dict_pass_string_crash": 31, "of_dict_pass_string_clear": 5, "of_dict_pass_string_ok": 3, "of_dict_pass_string_double": 1, "of_dict_pass_name_raw": 40, "cases": [ { "id": "C1_dict_args", "short": "dict_args", "title": "Dict arguments (baseline)", "kind": "render" }, { "id": "C2_string_args", "short": "string_args", "title": "OpenAI-style string arguments", "kind": "render" }, { "id": "C3_empty_args", "short": "empty_args", "title": "Empty dict arguments", "kind": "render" }, { "id": "C4_empty_string_args", "short": "empty_string_args", "title": "Empty string arguments", "kind": "render" }, { "id": "C5_nested_args", "short": "nested_args", "title": "Nested dict arguments", "kind": "render" }, { "id": "C6_nested_string_args", "short": "nested_string_args", "title": "Nested string arguments", "kind": "render" }, { "id": "C7_unicode_args", "short": "unicode_args", "title": "Unicode dict arguments", "kind": "render" }, { "id": "C8_unicode_string_args", "short": "unicode_string_args", "title": "Unicode string arguments", "kind": "render" }, { "id": "C9_multi_tool_calls", "short": "multi_tool_calls", "title": "Multiple tool calls", "kind": "render" }, { "id": "C10_name_escape", "short": "name_escape", "title": "Tool-name escaping", "kind": "escape" } ], "caveat": "A crash or a clear TemplateError on string arguments is a chat-template spec-compatibility issue, not a model-quality bug. transformers documents tool-call arguments as a mapping. This bench still scores OpenAI-style JSON strings because that is what many harnesses, traces, and datasets actually feed apply_chat_template.", "by_case": { "dict_args": { "pass": 40, "fail_crash": 1 }, "string_args": { "fail_crash": 32, "fail_clear": 5, "pass": 3, "fail_double_encoded": 1 }, "empty_args": { "pass": 40, "fail_crash": 1 }, "empty_string_args": { "fail_crash": 32, "fail_clear": 2, "pass": 6, "fail_double_encoded": 1 }, "nested_args": { "pass": 40, "fail_crash": 1 }, "nested_string_args": { "fail_crash": 32, "fail_clear": 5, "pass": 3, "fail_double_encoded": 1 }, "unicode_args": { "pass": 40, "fail_crash": 1 }, "unicode_string_args": { "fail_crash": 32, "fail_clear": 5, "pass": 3, "fail_double_encoded": 1 }, "multi_tool_calls": { "pass": 40, "fail_crash": 1 }, "name_escape": { "fail_raw_name": 40, "fail_crash": 1 } }, "of_dict_pass_empty_args_pass": 40, "of_dict_pass_nested_args_pass": 40, "of_dict_pass_unicode_args_pass": 40, "of_dict_pass_multi_pass": 40, "of_dict_pass_empty_string_pass": 6, "of_dict_pass_empty_string_crash": 31, "of_dict_pass_empty_string_clear": 2, "of_dict_pass_empty_string_double": 1, "best_score": 9, "n_with_best": 3, "n_perfect": 0 }