Datasets:
ToolUse.1: benchmark corpus, harness and specification
Browse files- LICENSE +201 -0
- README.md +178 -0
- SPEC.md +189 -0
- data/manifest.json +26 -0
- data/tooluse1.jsonl +114 -0
- pyproject.toml +32 -0
- requirements.txt +7 -0
- tests/test_corpus.py +130 -0
- tests/test_grade.py +175 -0
- tests/test_harness.py +163 -0
- tests/test_parse.py +142 -0
- tooluse1/__init__.py +52 -0
- tooluse1/__main__.py +4 -0
- tooluse1/cases_build.py +431 -0
- tooluse1/grade.py +196 -0
- tooluse1/harness.py +340 -0
- tooluse1/parse.py +219 -0
- tooluse1/prompts.py +133 -0
- tooluse1/report.py +124 -0
- tooluse1/runners.py +396 -0
- tooluse1/schema.py +187 -0
- tooluse1/tools.py +615 -0
LICENSE
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Apache License
|
| 2 |
+
Version 2.0, January 2004
|
| 3 |
+
http://www.apache.org/licenses/
|
| 4 |
+
|
| 5 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 6 |
+
|
| 7 |
+
1. Definitions.
|
| 8 |
+
|
| 9 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 10 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 11 |
+
|
| 12 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 13 |
+
the copyright owner that is granting the License.
|
| 14 |
+
|
| 15 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 16 |
+
other entities that control, are controlled by, or are under common
|
| 17 |
+
control with that entity. For the purposes of this definition,
|
| 18 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 19 |
+
direction or management of such entity, whether by contract or
|
| 20 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 21 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 22 |
+
|
| 23 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 24 |
+
exercising permissions granted by this License.
|
| 25 |
+
|
| 26 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 27 |
+
including but not limited to software source code, documentation
|
| 28 |
+
source, and configuration files.
|
| 29 |
+
|
| 30 |
+
"Object" form shall mean any form resulting from mechanical
|
| 31 |
+
transformation or translation of a Source form, including but
|
| 32 |
+
not limited to compiled object code, generated documentation,
|
| 33 |
+
and conversions to other media types.
|
| 34 |
+
|
| 35 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 36 |
+
Object form, made available under the License, as indicated by a
|
| 37 |
+
copyright notice that is included in or attached to the work
|
| 38 |
+
(an example is provided in the Appendix below).
|
| 39 |
+
|
| 40 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 41 |
+
form, that is based on (or derived from) the Work and for which the
|
| 42 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 43 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 44 |
+
of this License, Derivative Works shall not include works that remain
|
| 45 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 46 |
+
the Work and Derivative Works thereof.
|
| 47 |
+
|
| 48 |
+
"Contribution" shall mean any work of authorship, including
|
| 49 |
+
the original version of the Work and any modifications or additions
|
| 50 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 51 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 52 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 53 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 54 |
+
means any form of electronic, verbal, or written communication sent
|
| 55 |
+
to the Licensor or its representatives, including but not limited to
|
| 56 |
+
communication on electronic mailing lists, source code control systems,
|
| 57 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 58 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 59 |
+
excluding communication that is conspicuously marked or otherwise
|
| 60 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 61 |
+
|
| 62 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 63 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 64 |
+
subsequently incorporated within the Work.
|
| 65 |
+
|
| 66 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 67 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 68 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 69 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 70 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 71 |
+
Work and such Derivative Works in Source or Object form.
|
| 72 |
+
|
| 73 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 74 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 75 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 76 |
+
(except as stated in this section) patent license to make, have made,
|
| 77 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 78 |
+
where such license applies only to those patent claims licensable
|
| 79 |
+
by such Contributor that are necessarily infringed by their
|
| 80 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 81 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 82 |
+
institute patent litigation against any entity (including a
|
| 83 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 84 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 85 |
+
or contributory patent infringement, then any patent licenses
|
| 86 |
+
granted to You under this License for that Work shall terminate
|
| 87 |
+
as of the date such litigation is filed.
|
| 88 |
+
|
| 89 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 90 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 91 |
+
modifications, and in Source or Object form, provided that You
|
| 92 |
+
meet the following conditions:
|
| 93 |
+
|
| 94 |
+
(a) You must give any other recipients of the Work or
|
| 95 |
+
Derivative Works a copy of this License; and
|
| 96 |
+
|
| 97 |
+
(b) You must cause any modified files to carry prominent notices
|
| 98 |
+
stating that You changed the files; and
|
| 99 |
+
|
| 100 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 101 |
+
that You distribute, all copyright, patent, trademark, and
|
| 102 |
+
attribution notices from the Source form of the Work,
|
| 103 |
+
excluding those notices that do not pertain to any part of
|
| 104 |
+
the Derivative Works; and
|
| 105 |
+
|
| 106 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 107 |
+
distribution, then any Derivative Works that You distribute must
|
| 108 |
+
include a readable copy of the attribution notices contained
|
| 109 |
+
within such NOTICE file, excluding those notices that do not
|
| 110 |
+
pertain to any part of the Derivative Works, in at least one
|
| 111 |
+
of the following places: within a NOTICE text file distributed
|
| 112 |
+
as part of the Derivative Works; within the Source form or
|
| 113 |
+
documentation, if provided along with the Derivative Works; or,
|
| 114 |
+
within a display generated by the Derivative Works, if and
|
| 115 |
+
wherever such third-party notices normally appear. The contents
|
| 116 |
+
of the NOTICE file are for informational purposes only and
|
| 117 |
+
do not modify the License. You may add Your own attribution
|
| 118 |
+
notices within Derivative Works that You distribute, alongside
|
| 119 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 120 |
+
that such additional attribution notices cannot be construed
|
| 121 |
+
as modifying the License.
|
| 122 |
+
|
| 123 |
+
You may add Your own copyright statement to Your modifications and
|
| 124 |
+
may provide additional or different license terms and conditions
|
| 125 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 126 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 127 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 128 |
+
the conditions stated in this License.
|
| 129 |
+
|
| 130 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 131 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 132 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 133 |
+
this License, without any additional terms or conditions.
|
| 134 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 135 |
+
the terms of any separate license agreement you may have executed
|
| 136 |
+
with Licensor regarding such Contributions.
|
| 137 |
+
|
| 138 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 139 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 140 |
+
except as required for reasonable and customary use in describing the
|
| 141 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 142 |
+
|
| 143 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 144 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 145 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 146 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 147 |
+
implied, including, without limitation, any warranties or conditions
|
| 148 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 149 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 150 |
+
appropriateness of using or redistributing the Work and assume any
|
| 151 |
+
risks associated with Your exercise of permissions under this License.
|
| 152 |
+
|
| 153 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 154 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 155 |
+
unless required by applicable law (such as deliberate and grossly
|
| 156 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 157 |
+
liable to You for damages, including any direct, indirect, special,
|
| 158 |
+
incidental, or consequential damages of any character arising as a
|
| 159 |
+
result of this License or out of the use or inability to use the
|
| 160 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 161 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 162 |
+
other commercial damages or losses), even if such Contributor
|
| 163 |
+
has been advised of the possibility of such damages.
|
| 164 |
+
|
| 165 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 166 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 167 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 168 |
+
or other liability obligations and/or rights consistent with this
|
| 169 |
+
License. However, in accepting such obligations, You may act only
|
| 170 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 171 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 172 |
+
defend, and hold each Contributor harmless for any liability
|
| 173 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 174 |
+
of your accepting any such warranty or additional liability.
|
| 175 |
+
|
| 176 |
+
END OF TERMS AND CONDITIONS
|
| 177 |
+
|
| 178 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 179 |
+
|
| 180 |
+
To apply the Apache License to your work, attach the following
|
| 181 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 182 |
+
replaced with your own identifying information. (Don't include
|
| 183 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 184 |
+
comment syntax for the file format. We also recommend that a
|
| 185 |
+
file or class name and description of purpose be included on the
|
| 186 |
+
same "printed page" as the copyright notice for easier
|
| 187 |
+
identification within third-party archives.
|
| 188 |
+
|
| 189 |
+
Copyright 2026 Hardcore AI
|
| 190 |
+
|
| 191 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 192 |
+
you may not use this file except in compliance with the License.
|
| 193 |
+
You may obtain a copy of the License at
|
| 194 |
+
|
| 195 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 196 |
+
|
| 197 |
+
Unless required by applicable law or agreed to in writing, software
|
| 198 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 199 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 200 |
+
See the License for the specific language governing permissions and
|
| 201 |
+
limitations under the License.
|
README.md
ADDED
|
@@ -0,0 +1,178 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
- ru
|
| 6 |
+
pretty_name: ToolUse.1
|
| 7 |
+
task_categories:
|
| 8 |
+
- text-generation
|
| 9 |
+
tags:
|
| 10 |
+
- tool-use
|
| 11 |
+
- function-calling
|
| 12 |
+
- agents
|
| 13 |
+
- evaluation
|
| 14 |
+
- benchmark
|
| 15 |
+
- hardcore-ai
|
| 16 |
+
size_categories:
|
| 17 |
+
- n<1K
|
| 18 |
+
configs:
|
| 19 |
+
- config_name: default
|
| 20 |
+
data_files:
|
| 21 |
+
- split: test
|
| 22 |
+
path: data/tooluse1.jsonl
|
| 23 |
+
---
|
| 24 |
+
|
| 25 |
+
# ToolUse.1
|
| 26 |
+
|
| 27 |
+
A tool-calling benchmark that scores correct invocation **and** correct abstention.
|
| 28 |
+
|
| 29 |
+
114 cases · 78 requiring a call · 36 requiring abstention or clarification · English-only · 9 tool schemas · 3 calling protocols (JSON, XML, plain-text) · deterministic grading, no LLM judge.
|
| 30 |
+
|
| 31 |
+
Most tool-use evaluations reward a model for emitting a well-formed call. ToolUse.1 also asks
|
| 32 |
+
whether it should have called at all: roughly a third of the corpus is prompts where the correct
|
| 33 |
+
answer is to say something useful and call nothing. A model that calls a tool for every prompt
|
| 34 |
+
and a model that never calls one score about the same.
|
| 35 |
+
|
| 36 |
+
## Composition
|
| 37 |
+
|
| 38 |
+
| Category | Cases | Category | Cases |
|
| 39 |
+
|---|---|---|---|
|
| 40 |
+
| calculator | 15 | negative | 30 |
|
| 41 |
+
| get_weather | 8 | clarification | 5 |
|
| 42 |
+
| translate | 8 | multi_step | 6 |
|
| 43 |
+
| convert_currency | 8 | web_search | 8 |
|
| 44 |
+
| convert_units | 8 | set_timer | 6 |
|
| 45 |
+
| save_note | 6 | get_time | 6 |
|
| 46 |
+
|
| 47 |
+
114 English cases. Difficulty is labelled per case (`easy` / `medium` / `hard`) and
|
| 48 |
+
reported as a breakdown.
|
| 49 |
+
|
| 50 |
+
| Group | What it tests |
|
| 51 |
+
|---|---|
|
| 52 |
+
| `calculator`, `get_weather`, `translate`, `convert_currency`, `convert_units`, `set_timer`, `save_note`, `get_time`, `web_search` | Single call, correct tool, semantically correct arguments. Most cases ship all nine tools, so selection is tested against distractors. |
|
| 53 |
+
| `negative` | No call is correct. Knowledge questions, absent capabilities, explicit no-tool instructions, and a quoted call the user asked to be repeated. |
|
| 54 |
+
| `clarification` | A required argument is missing. Guessing fails; asking the user passes. |
|
| 55 |
+
| `multi_step` | History contains prior tool calls and results; the follow-up requires the next call, or a comparison answered without any call. |
|
| 56 |
+
|
| 57 |
+
## Schema
|
| 58 |
+
|
| 59 |
+
```json
|
| 60 |
+
{
|
| 61 |
+
"id": "weather-en-001",
|
| 62 |
+
"lang": "en",
|
| 63 |
+
"category": "get_weather",
|
| 64 |
+
"difficulty": "easy",
|
| 65 |
+
"tools": ["calculator", "get_weather", "..."],
|
| 66 |
+
"user": "What's the weather in Berlin?",
|
| 67 |
+
"expect": {"kind": "call", "name": "get_weather", "args": {"city": "Berlin"}}
|
| 68 |
+
}
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
`expect.kind` is `call` (with `name` and canonical `args`) or `no_call` (optionally with
|
| 72 |
+
`answer_contains_all` / `answer_contains_any`, case-insensitive substring groups). Multi-step
|
| 73 |
+
cases add a `history` array of `user` / `assistant` (with `tool_name`, `tool_args`) / `tool`
|
| 74 |
+
messages.
|
| 75 |
+
|
| 76 |
+
## Metrics
|
| 77 |
+
|
| 78 |
+
| Metric | Definition |
|
| 79 |
+
|---|---|
|
| 80 |
+
| **Score** | passed cases / all cases, with a Wilson 95% interval |
|
| 81 |
+
| Correct call when a tool applies | passing among call cases |
|
| 82 |
+
| Correct abstention | passing among no-call cases |
|
| 83 |
+
| **Spurious-call rate** | no-call cases where any call was emitted (lower is better) |
|
| 84 |
+
| Tool selected correctly | call cases with the right tool name |
|
| 85 |
+
| Arguments exactly correct | call cases with every expected argument matching |
|
| 86 |
+
| Malformed output rate | call attempts that failed to parse |
|
| 87 |
+
|
| 88 |
+
Arguments are compared **semantically**, not as strings: `expression` is evaluated as
|
| 89 |
+
arithmetic (`45+12` equals `45 + 12`), numbers compare with a `1e-6` tolerance, and cities,
|
| 90 |
+
languages, currencies, units and timezones go through alias tables (`СПб` → `saint petersburg`,
|
| 91 |
+
`английский` → `en`, `$` → `USD`, `Moscow` → `Europe/Moscow`). Tool names are alias tolerant.
|
| 92 |
+
Extra arguments are tolerated and flagged.
|
| 93 |
+
|
| 94 |
+
## Evaluation harness
|
| 95 |
+
|
| 96 |
+
The harness is in this repository (`tooluse1/`), alongside the frozen corpus in
|
| 97 |
+
`data/tooluse1.jsonl` and the corpus digest in `data/manifest.json`.
|
| 98 |
+
|
| 99 |
+
```bash
|
| 100 |
+
git clone https://huggingface.co/datasets/HardcoreAI/ToolUse.1
|
| 101 |
+
cd ToolUse.1
|
| 102 |
+
pip install -r requirements.txt
|
| 103 |
+
|
| 104 |
+
python -m tooluse1 selftest # grader sanity check, no model needed
|
| 105 |
+
python smoke.py --model HardcoreAI/Nova-1.1-0.8B --limit 8 --batch-size 8
|
| 106 |
+
python -m tooluse1 run --model Qwen/Qwen3.5-0.8B --backend transformers --protocol all --batch-size 8 --out results/
|
| 107 |
+
python -m tooluse1 run --model my-model --backend openai --base-url http://localhost:8000/v1 --protocol json
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
Backends: local `transformers` checkpoints, any OpenAI-compatible `/chat/completions` endpoint
|
| 111 |
+
(vLLM, llama.cpp server, Ollama, OpenRouter), or a scripted file for ablations. Local runs take
|
| 112 |
+
`--batch-size` (batch the evaluations into one padded `generate` call; 8–16 saturates a small
|
| 113 |
+
GPU) and `--dtype` (`float16` where bfloat16 is unavailable, e.g. a T4 — and report it).
|
| 114 |
+
|
| 115 |
+
Loading is verified, not assumed: the harness reads back the loading report and refuses an
|
| 116 |
+
architecture whose weights do not match the checkpoint, because `from_pretrained` only warns in
|
| 117 |
+
that case and hands you a randomly initialised model that still writes plausible text.
|
| 118 |
+
|
| 119 |
+
Reproduction settings: greedy decoding, `max_new_tokens=256`, the harness system prompt
|
| 120 |
+
unmodified, no few-shot examples, reasoning mode disabled where the runtime supports it. Use the
|
| 121 |
+
same protocol set when comparing numbers.
|
| 122 |
+
|
| 123 |
+
## Leaderboard
|
| 124 |
+
|
| 125 |
+
Results are listed on the [ToolUse.1 leaderboard](https://huggingface.co/spaces/HardcoreAI/ToolUse.1).
|
| 126 |
+
To submit, run the harness, keep the JSON it writes, and open a discussion here with the model
|
| 127 |
+
id, protocols, case count, decoding configuration and the raw results file.
|
| 128 |
+
|
| 129 |
+
The leaderboard seeds itself with harness self-checks only (an oracle runner that reads the
|
| 130 |
+
expectations, which must score exactly 1.000, and a null runner that never calls a tool). Those
|
| 131 |
+
rows validate the grader — they are not model results, and they are labelled as such.
|
| 132 |
+
|
| 133 |
+
## Contamination
|
| 134 |
+
|
| 135 |
+
The corpus is fully public. A model trained on these cases will overfit them. Scores should be
|
| 136 |
+
read as an upper bound for any model whose authors had access to this repository; models
|
| 137 |
+
evaluated before publication are the fair comparison. A held-out private split is planned for
|
| 138 |
+
the 2.x line.
|
| 139 |
+
|
| 140 |
+
## Versioning
|
| 141 |
+
|
| 142 |
+
Any change to a case bumps the minor version and changes the corpus digest in
|
| 143 |
+
`data/manifest.json`, which breaks comparability with earlier scores. Grading changes bump the
|
| 144 |
+
major version. Always report the digest with a number.
|
| 145 |
+
|
| 146 |
+
## Files
|
| 147 |
+
|
| 148 |
+
```
|
| 149 |
+
data/tooluse1.jsonl 114 cases, one JSON object per line
|
| 150 |
+
data/manifest.json counts, sha256 of the frozen corpus, harness version
|
| 151 |
+
SPEC.md full specification: grading rules, limitations, contamination policy
|
| 152 |
+
smoke.py checkpoint loading check plus a throughput measurement
|
| 153 |
+
tooluse1/ harness (cases, tools, protocols, parser, grader, runners, CLI)
|
| 154 |
+
tests/ tests covering parser robustness, grading semantics, corpus invariants
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
## Limitations
|
| 158 |
+
|
| 159 |
+
See `SPEC.md` §6. The short version: substring-based answer constraints, wording-sensitive
|
| 160 |
+
clarification expectations, synthetic tool results in multi-step cases, no private split yet, and
|
| 161 |
+
one documented ambiguity in the plain-text protocol where a call quoted on its own line is
|
| 162 |
+
indistinguishable from a real call.
|
| 163 |
+
|
| 164 |
+
## Citation
|
| 165 |
+
|
| 166 |
+
```bibtex
|
| 167 |
+
@misc{hardcoreai_tooluse1,
|
| 168 |
+
title = {ToolUse.1: a tool-calling benchmark with abstention},
|
| 169 |
+
author = {Hardcore AI},
|
| 170 |
+
year = {2026},
|
| 171 |
+
url = {https://huggingface.co/datasets/HardcoreAI/ToolUse.1}
|
| 172 |
+
}
|
| 173 |
+
```
|
| 174 |
+
|
| 175 |
+
## License
|
| 176 |
+
|
| 177 |
+
Apache-2.0. The corpus may be redistributed and used for training; see the contamination note
|
| 178 |
+
above before comparing numbers.
|
SPEC.md
ADDED
|
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ToolUse.1 — Specification
|
| 2 |
+
|
| 3 |
+
**Version:** 1.0.0 · **Corpus digest:** see `data/manifest.json` · **License:** Apache-2.0
|
| 4 |
+
|
| 5 |
+
ToolUse.1 measures whether a chat model invokes tools **correctly and only when it should**.
|
| 6 |
+
It is deliberately not a "can the model produce a JSON blob" test. Three things are scored
|
| 7 |
+
separately and reported side by side:
|
| 8 |
+
|
| 9 |
+
1. **Correct invocation** — the right tool, with the right arguments, when a tool applies.
|
| 10 |
+
2. **Correct abstention** — no call at all when no tool applies.
|
| 11 |
+
3. **Correct clarification** — a question back to the user when a required argument is missing
|
| 12 |
+
and cannot be inferred.
|
| 13 |
+
|
| 14 |
+
A model that calls a tool for every prompt scores the same as one that never calls anything:
|
| 15 |
+
both fail roughly half the corpus.
|
| 16 |
+
|
| 17 |
+
## 1. Corpus
|
| 18 |
+
|
| 19 |
+
188 authored cases, 128 requiring a call and 60 requiring abstention or clarification.
|
| 20 |
+
|
| 21 |
+
| Category | Cases | Category | Cases |
|
| 22 |
+
|---|---|---|---|
|
| 23 |
+
| calculator | 15 | negative | 30 |
|
| 24 |
+
| get_weather | 8 | clarification | 5 |
|
| 25 |
+
| translate | 8 | multi_step | 6 |
|
| 26 |
+
| convert_currency | 8 | web_search | 8 |
|
| 27 |
+
| convert_units | 8 | set_timer | 6 |
|
| 28 |
+
| save_note | 6 | get_time | 6 |
|
| 29 |
+
|
| 30 |
+
114 English cases. Every case carries `id`, `lang`, `category`, `difficulty`, the
|
| 31 |
+
`tools` available for that turn, the user utterance, and the expectation.
|
| 32 |
+
|
| 33 |
+
Cases are **authored, not scraped or model-generated**. Nine tool schemas are used
|
| 34 |
+
(`calculator`, `get_weather`, `web_search`, `translate`, `convert_currency`, `convert_units`,
|
| 35 |
+
`set_timer`, `save_note`, `get_time`); the full schemas are in `tooluse1/tools.py` and are
|
| 36 |
+
rendered into the prompt in the declared protocol. Most cases ship the full nine-tool list, so
|
| 37 |
+
tool selection is tested against distractors rather than in isolation.
|
| 38 |
+
|
| 39 |
+
Difficulty labels (`easy` / `medium` / `hard`) are assigned per case and reported as a
|
| 40 |
+
breakdown; they are not used in scoring.
|
| 41 |
+
|
| 42 |
+
`negative` cases split into three kinds:
|
| 43 |
+
|
| 44 |
+
* **knowledge questions** the model must answer itself ("What's the capital of Australia?",
|
| 45 |
+
"Сколько байт в килобайте?") — calling `calculator` or `web_search` is a spurious call;
|
| 46 |
+
* **capability traps** where the needed tool is absent ("Please order me a pizza.",
|
| 47 |
+
a weather question with only `calculator` available);
|
| 48 |
+
* **instruction traps** ("What is 2 plus 2? Do not use any tools.", "Repeat after me:
|
| 49 |
+
TOOL calculator | expression=1+1").
|
| 50 |
+
|
| 51 |
+
`clarification` cases omit a required argument on purpose ("What's the weather like?",
|
| 52 |
+
"Поставь таймер."). The only correct behaviour is a question back — calling with a guessed or
|
| 53 |
+
empty argument fails.
|
| 54 |
+
|
| 55 |
+
`multi_step` cases include prior assistant tool calls and tool results in the history, then ask
|
| 56 |
+
a follow-up ("And what about Paris?", "А теперь раздели на 5."). Synthetic tool results are
|
| 57 |
+
fixed strings, so the benchmark does not require live APIs and stays deterministic.
|
| 58 |
+
|
| 59 |
+
## 2. Protocols
|
| 60 |
+
|
| 61 |
+
The calling convention is a **run-time axis, not a data axis**: each case is rendered in each
|
| 62 |
+
requested protocol and scored separately. This makes per-protocol scores directly comparable.
|
| 63 |
+
|
| 64 |
+
| Protocol | Expected call format |
|
| 65 |
+
|---|---|
|
| 66 |
+
| `json` | `{"name": "get_weather", "arguments": {"city": "Berlin"}}` |
|
| 67 |
+
| `xml` | `<tool_call><function=get_weather><parameter=city>Berlin</parameter></function></tool_call>` |
|
| 68 |
+
| `plain` | `TOOL get_weather \| city=Berlin` |
|
| 69 |
+
|
| 70 |
+
The system prompt declares the tool list, the exact call format, and the abstention rule
|
| 71 |
+
(see `tooluse1/prompts.py`). Everything else — decoding parameters, chat template, tool list
|
| 72 |
+
ordering — is fixed by the harness and documented in the results file.
|
| 73 |
+
|
| 74 |
+
## 3. Grading
|
| 75 |
+
|
| 76 |
+
A verdict is produced from the raw model output by `tooluse1/parse.py` and `tooluse1/grade.py`.
|
| 77 |
+
Nothing is judged by a model; grading is deterministic and reproducible.
|
| 78 |
+
|
| 79 |
+
### 3.1 Extracting a call
|
| 80 |
+
|
| 81 |
+
Extractors are tried in priority order and the first that yields anything wins: XML call
|
| 82 |
+
blocks, JSON objects (bare, fenced, embedded in prose, OpenAI `tool_calls` shape, `function` +
|
| 83 |
+
string `arguments`), the plain-text `TOOL name | k=v` line, and a bare argument object whose
|
| 84 |
+
keys uniquely identify one tool. Qwen-style `thinking` blocks and `｜end▁of▁thinking｜` markers
|
| 85 |
+
are stripped before extraction.
|
| 86 |
+
|
| 87 |
+
A call case requires **exactly one** extracted call. Two calls fail the case.
|
| 88 |
+
|
| 89 |
+
### 3.2 Comparing arguments
|
| 90 |
+
|
| 91 |
+
Comparison is semantic, per tool. Each tool declares normalisers for its parameters:
|
| 92 |
+
|
| 93 |
+
* `expression` — evaluated as arithmetic (safe AST walk; no names other than a fixed math
|
| 94 |
+
whitelist). `45+12`, `45 + 12` and `(45+12)` all normalise to `57`.
|
| 95 |
+
* numeric arguments — compared with `rel_tol=1e-6`, so `300` equals `300.0` and `100` equals `"100"`.
|
| 96 |
+
* cities, languages, currencies, units, timezones — alias tables in `tooluse1/tools.py`
|
| 97 |
+
(`СПб` → `saint petersburg`, `английский` → `en`, `$` → `USD`, `градусов` → `celsius`,
|
| 98 |
+
`Moscow` → `Europe/Moscow`).
|
| 99 |
+
* free text — Unicode-normalised, whitespace-collapsed, case-insensitive.
|
| 100 |
+
|
| 101 |
+
`args_ok` requires every expected parameter to be present and equal. Extra parameters are
|
| 102 |
+
**tolerated** (flagged as `extra`) so that a model returning the optional `unit` is not punished.
|
| 103 |
+
|
| 104 |
+
Tool names are alias-tolerant (`calc` → `calculator`, `weather` → `get_weather`), which is how a
|
| 105 |
+
lenient runtime would resolve them.
|
| 106 |
+
|
| 107 |
+
### 3.3 Abstention
|
| 108 |
+
|
| 109 |
+
A no-call case passes when no call is extracted. If the case declares answer constraints
|
| 110 |
+
(`answer_contains_all`, `answer_contains_any`), the prose remainder after removing call
|
| 111 |
+
payloads and thinking must satisfy them, case-insensitively. Unconstrained no-call cases
|
| 112 |
+
("Tell me a joke.") accept any answer.
|
| 113 |
+
|
| 114 |
+
### 3.4 Metrics
|
| 115 |
+
|
| 116 |
+
| Metric | Definition |
|
| 117 |
+
|---|---|
|
| 118 |
+
| **Score** | passed cases / all cases |
|
| 119 |
+
| Correct call when a tool applies | passing cases among call cases |
|
| 120 |
+
| Correct abstention | passing cases among no-call cases |
|
| 121 |
+
| **Spurious-call rate** | no-call cases where *any* call was emitted |
|
| 122 |
+
| Tool selected correctly | call cases where the tool name matched |
|
| 123 |
+
| Arguments exactly correct | call cases where every expected argument matched |
|
| 124 |
+
| Answer constraints satisfied | no-call cases with constraints where prose matched |
|
| 125 |
+
| Malformed output rate | outputs that look like a call attempt but could not be parsed |
|
| 126 |
+
|
| 127 |
+
`Score` is the headline number and the only one used for ranking. The 95% confidence interval
|
| 128 |
+
is the Wilson score interval. Every metric is additionally reported per protocol, language,
|
| 129 |
+
category and difficulty.
|
| 130 |
+
|
| 131 |
+
## 4. Running the benchmark
|
| 132 |
+
|
| 133 |
+
```bash
|
| 134 |
+
pip install -r requirements.txt
|
| 135 |
+
python -m tooluse1 build # regenerate data/tooluse1.jsonl from source
|
| 136 |
+
python -m tooluse1 selftest # verify the grader against known-good runners
|
| 137 |
+
python -m tooluse1 run --model <hf-id> --backend transformers --protocol all
|
| 138 |
+
python -m tooluse1 run --model <name> --backend openai --base-url http://localhost:8000/v1
|
| 139 |
+
```
|
| 140 |
+
|
| 141 |
+
Deterministic settings anyone reproducing the numbers should use: greedy decoding
|
| 142 |
+
(`do_sample=False`), `max_new_tokens=256`, the harness system prompt unmodified, all three
|
| 143 |
+
protocols, no few-shot examples, thinking/reasoning mode disabled by the runtime where the
|
| 144 |
+
model supports it. `--resume` continues an interrupted run; `selftest` proves the grader agrees
|
| 145 |
+
with the corpus before any model is loaded (oracle-style runners must score exactly 1.000).
|
| 146 |
+
|
| 147 |
+
Local runs may pass `--batch-size N` to evaluate N prompts in one padded `generate` call. Greedy
|
| 148 |
+
decoding is unchanged by batching in theory but not bit-identical in practice (kernel reduction
|
| 149 |
+
orders differ), so a borderline case can flip; report the batch size with the number, and use
|
| 150 |
+
the same batch size when comparing two checkpoints. The measured reference on an RTX 4050
|
| 151 |
+
Laptop, bfloat16, batch 8: 0.66 s per evaluation, i.e. ~6 minutes for all 564.
|
| 152 |
+
|
| 153 |
+
The harness verifies how a checkpoint loaded: it inspects the loading report from
|
| 154 |
+
`from_pretrained` and rejects an architecture whose keys do not match, rather than silently
|
| 155 |
+
scoring a partially randomly initialised model. Nova-1.1 carries text-only keys plus a vision
|
| 156 |
+
tower, so the multimodal wrapper loads "successfully" and generates noise — the check is what
|
| 157 |
+
keeps that out of the results.
|
| 158 |
+
|
| 159 |
+
## 5. Contamination and versioning
|
| 160 |
+
|
| 161 |
+
The corpus is public in its entirety. That is a deliberate trade: it makes results checkable,
|
| 162 |
+
and it means a model trained on ToolUse.1 cases will overfit it. Treat scores as an upper
|
| 163 |
+
bound for any model whose authors had access to this repository, and prefer models evaluated
|
| 164 |
+
before the corpus was published. A held-out private split is planned for 2.x.
|
| 165 |
+
|
| 166 |
+
Versioning follows SemVer. Adding or editing a case bumps the minor version and changes the
|
| 167 |
+
corpus digest, which invalidates comparability of past scores; changing grading semantics bumps
|
| 168 |
+
the major version. The digest is recorded in `data/manifest.json` and should accompany any
|
| 169 |
+
reported number.
|
| 170 |
+
|
| 171 |
+
## 6. Known limitations
|
| 172 |
+
|
| 173 |
+
* The plain-text protocol cannot distinguish a call from a quoted `TOOL ...` string placed on
|
| 174 |
+
its own line. One trap case relies on this; it is a property of the protocol, not a bug.
|
| 175 |
+
* Answer constraints are substring checks, so an answer may satisfy a constraint for the wrong
|
| 176 |
+
reason. They cover facts with unambiguous wording (Capitals, years, 1024, 57) and are never
|
| 177 |
+
the sole basis of a case that has no call component.
|
| 178 |
+
* Clarification cases necessarily accept a range of phrasings; English phrasings
|
| 179 |
+
are enumerated in the expectations.
|
| 180 |
+
* Multi-step cases replay *synthetic* tool results. They test the sequencing decision, not the
|
| 181 |
+
model's ability to consume a real API response.
|
| 182 |
+
* Arguments that a runtime would fill in (dates, defaults) are out of scope; every required
|
| 183 |
+
argument is stated in the utterance or present in the history.
|
| 184 |
+
|
| 185 |
+
## 7. Submitting results
|
| 186 |
+
|
| 187 |
+
Run the harness, keep the generated JSON, and open a discussion on the dataset repository with
|
| 188 |
+
the model id, the protocol(s), the case count, the decoding configuration and the raw results
|
| 189 |
+
file. Results are listed on the leaderboard Space with a link to the raw file.
|
data/manifest.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"benchmark": "ToolUse.1",
|
| 3 |
+
"version": "1.0.0",
|
| 4 |
+
"cases": 114,
|
| 5 |
+
"call_cases": 78,
|
| 6 |
+
"no_call_cases": 36,
|
| 7 |
+
"by_language": {
|
| 8 |
+
"en": 114
|
| 9 |
+
},
|
| 10 |
+
"by_category": {
|
| 11 |
+
"calculator": 15,
|
| 12 |
+
"get_weather": 8,
|
| 13 |
+
"web_search": 8,
|
| 14 |
+
"translate": 8,
|
| 15 |
+
"convert_currency": 8,
|
| 16 |
+
"convert_units": 8,
|
| 17 |
+
"set_timer": 6,
|
| 18 |
+
"save_note": 6,
|
| 19 |
+
"get_time": 6,
|
| 20 |
+
"negative": 30,
|
| 21 |
+
"clarification": 5,
|
| 22 |
+
"multi_step": 6
|
| 23 |
+
},
|
| 24 |
+
"sha256": "17a6ef84c22648d83add6f56a03b24f896420837f4e69039c22d61cf8ad2c3e1",
|
| 25 |
+
"source": "C:\\models\\code\\tooluse1\\data\\tooluse1.jsonl"
|
| 26 |
+
}
|
data/tooluse1.jsonl
ADDED
|
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "45 + 12"}, "kind": "call", "name": "calculator"}, "id": "calc-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is 45 plus 12?"}
|
| 2 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "128 * 7"}, "kind": "call", "name": "calculator"}, "id": "calc-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Calculate 128 * 7 for me."}
|
| 3 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "0.15 * 240"}, "kind": "call", "name": "calculator"}, "id": "calc-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's 15% of 240?"}
|
| 4 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "(18 + 24) / 6"}, "kind": "call", "name": "calculator"}, "id": "calc-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Compute (18 + 24) / 6."}
|
| 5 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "2 ** 10"}, "kind": "call", "name": "calculator"}, "id": "calc-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How much is 2 to the power of 10?"}
|
| 6 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "sqrt(144)"}, "kind": "call", "name": "calculator"}, "id": "calc-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is the square root of 144?"}
|
| 7 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "0.1 + 0.2"}, "kind": "call", "name": "calculator"}, "id": "calc-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Add 0.1 and 0.2."}
|
| 8 |
+
{"category": "calculator", "difficulty": "easy", "expect": {"args": {"expression": "17 * 3 - 5"}, "kind": "call", "name": "calculator"}, "id": "calc-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's 17 times 3 minus 5?"}
|
| 9 |
+
{"category": "calculator", "difficulty": "medium", "expect": {"args": {"expression": "849 * 3"}, "kind": "call", "name": "calculator"}, "id": "calc-en-009", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "A shirt costs 849 and I want three of them. What's the total?"}
|
| 10 |
+
{"category": "calculator", "difficulty": "medium", "expect": {"args": {"expression": "(1000 / 8) ** 2"}, "kind": "call", "name": "calculator"}, "id": "calc-en-010", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's 1000 divided by 8, then squared?"}
|
| 11 |
+
{"category": "calculator", "difficulty": "medium", "expect": {"args": {"expression": "round(187.456)"}, "kind": "call", "name": "calculator"}, "id": "calc-en-011", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Round 187.456 to the nearest integer."}
|
| 12 |
+
{"category": "calculator", "difficulty": "medium", "expect": {"args": {"expression": "155 % 7"}, "kind": "call", "name": "calculator"}, "id": "calc-en-012", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the remainder when 155 is divided by 7?"}
|
| 13 |
+
{"category": "calculator", "difficulty": "hard", "expect": {"args": {"expression": "3.5 ** 4"}, "kind": "call", "name": "calculator"}, "id": "calc-en-013", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How much is 3.5 raised to the power of 4?"}
|
| 14 |
+
{"category": "calculator", "difficulty": "hard", "expect": {"args": {"expression": "log10(1000)"}, "kind": "call", "name": "calculator"}, "id": "calc-en-014", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is log10 of 1000?"}
|
| 15 |
+
{"category": "calculator", "difficulty": "hard", "expect": {"args": {"expression": "(14 + 27 + 40) / 3"}, "kind": "call", "name": "calculator"}, "id": "calc-en-015", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the average of 14, 27 and 40?"}
|
| 16 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "Berlin"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather in Berlin?"}
|
| 17 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "Tokyo"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Is it raining in Tokyo right now?"}
|
| 18 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "Moscow"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Tell me the current temperature in Moscow."}
|
| 19 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "New York"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather like in New York today?"}
|
| 20 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "Saint Petersburg"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather in St. Petersburg?"}
|
| 21 |
+
{"category": "get_weather", "difficulty": "easy", "expect": {"args": {"city": "Chicago", "unit": "fahrenheit"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather in Chicago in Fahrenheit?"}
|
| 22 |
+
{"category": "get_weather", "difficulty": "medium", "expect": {"args": {"city": "Dubai", "unit": "celsius"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Current temperature in Dubai in Celsius?"}
|
| 23 |
+
{"category": "get_weather", "difficulty": "hard", "expect": {"args": {"city": "London"}, "kind": "call", "name": "get_weather"}, "id": "weather-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Do I need an umbrella in London this afternoon?"}
|
| 24 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "latest stable Rust version"}, "kind": "call", "name": "web_search"}, "id": "search-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Search the web for the latest stable version of Rust."}
|
| 25 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "CEO of NVIDIA"}, "kind": "call", "name": "web_search"}, "id": "search-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Look up who the CEO of NVIDIA is."}
|
| 26 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "RTX 4050 laptop price"}, "kind": "call", "name": "web_search"}, "id": "search-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Find the current price of an RTX 4050 laptop."}
|
| 27 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "Framework 13 laptop review"}, "kind": "call", "name": "web_search"}, "id": "search-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Search for reviews of the Framework 13 laptop."}
|
| 28 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "population of Almaty"}, "kind": "call", "name": "web_search"}, "id": "search-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Look up the population of Almaty."}
|
| 29 |
+
{"category": "web_search", "difficulty": "easy", "expect": {"args": {"query": "next Formula 1 race"}, "kind": "call", "name": "web_search"}, "id": "search-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Find out when the next Formula 1 race is."}
|
| 30 |
+
{"category": "web_search", "difficulty": "medium", "expect": {"args": {"query": "Python 3.13 release notes"}, "kind": "call", "name": "web_search"}, "id": "search-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Search the web for Python 3.13 release notes."}
|
| 31 |
+
{"category": "web_search", "difficulty": "hard", "expect": {"args": {"query": "Nova-1.1-0.8B GGUF"}, "kind": "call", "name": "web_search"}, "id": "search-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Look up whether the Nova-1.1-0.8B model card mentions GGUF."}
|
| 32 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "fr", "text": "good morning"}, "kind": "call", "name": "translate"}, "id": "translate-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'good morning' to French."}
|
| 33 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "es", "text": "thank you"}, "kind": "call", "name": "translate"}, "id": "translate-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'thank you' into Spanish."}
|
| 34 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "de", "text": "how are you"}, "kind": "call", "name": "translate"}, "id": "translate-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How do you say 'how are you' in German?"}
|
| 35 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "it", "text": "I would like a coffee"}, "kind": "call", "name": "translate"}, "id": "translate-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'I would like a coffee' to Italian."}
|
| 36 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "ja", "text": "hello"}, "kind": "call", "name": "translate"}, "id": "translate-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'hello' to Japanese."}
|
| 37 |
+
{"category": "translate", "difficulty": "easy", "expect": {"args": {"target_language": "ru", "text": "where is the station"}, "kind": "call", "name": "translate"}, "id": "translate-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'where is the station' into Russian."}
|
| 38 |
+
{"category": "translate", "difficulty": "medium", "expect": {"args": {"target_language": "pt", "text": "the book is on the table"}, "kind": "call", "name": "translate"}, "id": "translate-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'the book is on the table' into Portuguese."}
|
| 39 |
+
{"category": "translate", "difficulty": "medium", "expect": {"args": {"target_language": "tr", "text": "see you tomorrow"}, "kind": "call", "name": "translate"}, "id": "translate-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Translate 'see you tomorrow' to Turkish."}
|
| 40 |
+
{"category": "convert_currency", "difficulty": "easy", "expect": {"args": {"amount": 100, "from_currency": "USD", "to_currency": "EUR"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 100 USD to EUR."}
|
| 41 |
+
{"category": "convert_currency", "difficulty": "easy", "expect": {"args": {"amount": 500, "from_currency": "USD", "to_currency": "RUB"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How many rubles is 500 dollars?"}
|
| 42 |
+
{"category": "convert_currency", "difficulty": "easy", "expect": {"args": {"amount": 250, "from_currency": "EUR", "to_currency": "GBP"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 250 euros to pounds."}
|
| 43 |
+
{"category": "convert_currency", "difficulty": "easy", "expect": {"args": {"amount": 1000, "from_currency": "RUB", "to_currency": "KZT"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's 1000 rubles in tenge?"}
|
| 44 |
+
{"category": "convert_currency", "difficulty": "medium", "expect": {"args": {"amount": 75, "from_currency": "GBP", "to_currency": "JPY"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 75 GBP to JPY."}
|
| 45 |
+
{"category": "convert_currency", "difficulty": "easy", "expect": {"args": {"amount": 50, "from_currency": "EUR", "to_currency": "USD"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How much is 50 EUR in USD?"}
|
| 46 |
+
{"category": "convert_currency", "difficulty": "medium", "expect": {"args": {"amount": 30, "from_currency": "CAD", "to_currency": "USD"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 30 Canadian dollars to US dollars."}
|
| 47 |
+
{"category": "convert_currency", "difficulty": "hard", "expect": {"args": {"amount": 20000, "from_currency": "KZT", "to_currency": "RUB"}, "kind": "call", "name": "convert_currency"}, "id": "currency-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Exchange 20000 tenge for rubles."}
|
| 48 |
+
{"category": "convert_units", "difficulty": "easy", "expect": {"args": {"from_unit": "km", "to_unit": "mi", "value": 5}, "kind": "call", "name": "convert_units"}, "id": "units-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 5 km to miles."}
|
| 49 |
+
{"category": "convert_units", "difficulty": "easy", "expect": {"args": {"from_unit": "kg", "to_unit": "lb", "value": 70}, "kind": "call", "name": "convert_units"}, "id": "units-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How many pounds is 70 kg?"}
|
| 50 |
+
{"category": "convert_units", "difficulty": "easy", "expect": {"args": {"from_unit": "cm", "to_unit": "ft", "value": 180}, "kind": "call", "name": "convert_units"}, "id": "units-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 180 cm to feet."}
|
| 51 |
+
{"category": "convert_units", "difficulty": "easy", "expect": {"args": {"from_unit": "fahrenheit", "to_unit": "celsius", "value": 100}, "kind": "call", "name": "convert_units"}, "id": "units-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 100 Fahrenheit to Celsius."}
|
| 52 |
+
{"category": "convert_units", "difficulty": "medium", "expect": {"args": {"from_unit": "mi", "to_unit": "km", "value": 26.2}, "kind": "call", "name": "convert_units"}, "id": "units-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How many kilometers is 26.2 miles?"}
|
| 53 |
+
{"category": "convert_units", "difficulty": "medium", "expect": {"args": {"from_unit": "l", "to_unit": "gal", "value": 2}, "kind": "call", "name": "convert_units"}, "id": "units-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 2 liters to gallons."}
|
| 54 |
+
{"category": "convert_units", "difficulty": "medium", "expect": {"args": {"from_unit": "ml", "to_unit": "l", "value": 500}, "kind": "call", "name": "convert_units"}, "id": "units-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 500 milliliters to liters."}
|
| 55 |
+
{"category": "convert_units", "difficulty": "hard", "expect": {"args": {"from_unit": "min", "to_unit": "h", "value": 90}, "kind": "call", "name": "convert_units"}, "id": "units-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert 90 minutes to hours."}
|
| 56 |
+
{"category": "set_timer", "difficulty": "easy", "expect": {"args": {"seconds": 300}, "kind": "call", "name": "set_timer"}, "id": "timer-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Set a 5 minute timer."}
|
| 57 |
+
{"category": "set_timer", "difficulty": "easy", "expect": {"args": {"seconds": 600}, "kind": "call", "name": "set_timer"}, "id": "timer-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Remind me in 10 minutes."}
|
| 58 |
+
{"category": "set_timer", "difficulty": "easy", "expect": {"args": {"seconds": 90}, "kind": "call", "name": "set_timer"}, "id": "timer-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Set a timer for 90 seconds."}
|
| 59 |
+
{"category": "set_timer", "difficulty": "easy", "expect": {"args": {"seconds": 3600}, "kind": "call", "name": "set_timer"}, "id": "timer-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Start a one hour timer."}
|
| 60 |
+
{"category": "set_timer", "difficulty": "medium", "expect": {"args": {"seconds": 1500}, "kind": "call", "name": "set_timer"}, "id": "timer-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Put a 25 minute pomodoro timer on."}
|
| 61 |
+
{"category": "set_timer", "difficulty": "hard", "expect": {"args": {"seconds": 150}, "kind": "call", "name": "set_timer"}, "id": "timer-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Set a timer for two and a half minutes."}
|
| 62 |
+
{"category": "save_note", "difficulty": "easy", "expect": {"args": {"content": "Tuesday at 3 PM", "title": "Meeting"}, "kind": "call", "name": "save_note"}, "id": "note-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Save a note titled 'Meeting' with the text 'Tuesday at 3 PM'."}
|
| 63 |
+
{"category": "save_note", "difficulty": "easy", "expect": {"args": {"content": "buy milk", "title": "Groceries"}, "kind": "call", "name": "save_note"}, "id": "note-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Add a note called 'Groceries' that says buy milk."}
|
| 64 |
+
{"category": "save_note", "difficulty": "easy", "expect": {"args": {"content": "build a benchmark", "title": "Ideas"}, "kind": "call", "name": "save_note"}, "id": "note-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Save a note titled 'Ideas' with content 'build a benchmark'."}
|
| 65 |
+
{"category": "save_note", "difficulty": "easy", "expect": {"args": {"content": "backup at 02:00", "title": "Server"}, "kind": "call", "name": "save_note"}, "id": "note-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Save a note titled 'Server' with the text 'backup at 02:00'."}
|
| 66 |
+
{"category": "save_note", "difficulty": "medium", "expect": {"args": {"content": "stay hungry", "title": "Quote"}, "kind": "call", "name": "save_note"}, "id": "note-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Create a note called 'Quote' with the text 'stay hungry'."}
|
| 67 |
+
{"category": "save_note", "difficulty": "hard", "expect": {"args": {"content": "buy concert tickets", "title": "Tickets"}, "kind": "call", "name": "save_note"}, "id": "note-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Save a reminder note titled 'Tickets' with content 'buy concert tickets'."}
|
| 68 |
+
{"category": "get_time", "difficulty": "easy", "expect": {"args": {"timezone": "Asia/Tokyo"}, "kind": "call", "name": "get_time"}, "id": "time-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What time is it in Tokyo?"}
|
| 69 |
+
{"category": "get_time", "difficulty": "easy", "expect": {"args": {"timezone": "America/New_York"}, "kind": "call", "name": "get_time"}, "id": "time-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the current time in New York?"}
|
| 70 |
+
{"category": "get_time", "difficulty": "easy", "expect": {"args": {"timezone": "Europe/London"}, "kind": "call", "name": "get_time"}, "id": "time-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What time is it in London right now?"}
|
| 71 |
+
{"category": "get_time", "difficulty": "easy", "expect": {"args": {"timezone": "Europe/Moscow"}, "kind": "call", "name": "get_time"}, "id": "time-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What time is it in Moscow?"}
|
| 72 |
+
{"category": "get_time", "difficulty": "easy", "expect": {"args": {"timezone": "Europe/Berlin"}, "kind": "call", "name": "get_time"}, "id": "time-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What time is it in Berlin?"}
|
| 73 |
+
{"category": "get_time", "difficulty": "medium", "expect": {"args": {"timezone": "Australia/Sydney"}, "kind": "call", "name": "get_time"}, "id": "time-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the date and time in Sydney?"}
|
| 74 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["canberra"], "kind": "no_call"}, "id": "negative-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is the capital of Australia?"}
|
| 75 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_any": [["tolstoy", "leo tolstoy"]], "kind": "no_call"}, "id": "negative-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Who wrote 'War and Peace'?"}
|
| 76 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["hash"], "kind": "no_call"}, "id": "negative-en-003", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Explain what a hash function is."}
|
| 77 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["tcp", "udp"], "kind": "no_call"}, "id": "negative-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the difference between TCP and UDP?"}
|
| 78 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Tell me a joke."}
|
| 79 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-en-006", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's your name?"}
|
| 80 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["learning"], "kind": "no_call"}, "id": "negative-en-007", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is machine learning?"}
|
| 81 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-en-008", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Write a short poem about the sea."}
|
| 82 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-en-009", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Is Python or Rust a better choice for CLI tools?"}
|
| 83 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_any": [["not found", "not be found"]], "kind": "no_call"}, "id": "negative-en-010", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What does HTTP 404 mean?"}
|
| 84 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["recursion"], "kind": "no_call"}, "id": "negative-en-011", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Can you explain recursion?"}
|
| 85 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["sleep"], "kind": "no_call"}, "id": "negative-en-012", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Give me three tips for better sleep."}
|
| 86 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_any": [["leonardo", "da vinci", "vinci"]], "kind": "no_call"}, "id": "negative-en-013", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Who painted the Mona Lisa?"}
|
| 87 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["100"], "kind": "no_call"}, "id": "negative-en-014", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the boiling point of water at sea level?"}
|
| 88 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["api"], "kind": "no_call"}, "id": "negative-en-015", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Summarize what an API is in one sentence."}
|
| 89 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["1989"], "kind": "no_call"}, "id": "negative-en-016", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "In what year did the Berlin Wall fall?"}
|
| 90 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["ottawa"], "kind": "no_call"}, "id": "negative-en-017", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the capital of Canada?"}
|
| 91 |
+
{"category": "negative", "difficulty": "easy", "expect": {"answer_contains_all": ["list"], "kind": "no_call"}, "id": "negative-en-018", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Explain the difference between a list and a tuple in Python."}
|
| 92 |
+
{"category": "negative", "difficulty": "medium", "expect": {"answer_contains_all": ["57"], "kind": "no_call"}, "id": "negative-hard-en-001", "lang": "en", "notes": "calculator is not in the tool list; answering from knowledge is correct", "tools": ["get_weather"], "user": "What is 45 plus 12?"}
|
| 93 |
+
{"category": "negative", "difficulty": "medium", "expect": {"answer_contains_any": [["can't", "cannot", "don't have", "do not have", "unable", "no access", "weather tool"]], "kind": "no_call"}, "id": "negative-hard-en-002", "lang": "en", "notes": "no weather tool is available", "tools": ["calculator"], "user": "What's the weather in Paris?"}
|
| 94 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-hard-en-003", "lang": "en", "notes": "no ordering tool exists", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Please order me a pizza."}
|
| 95 |
+
{"category": "negative", "difficulty": "easy", "expect": {"kind": "no_call"}, "id": "negative-hard-en-004", "lang": "en", "notes": "no home-automation tool exists", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "Turn on the lights."}
|
| 96 |
+
{"category": "negative", "difficulty": "hard", "expect": {"answer_contains_all": ["1024"], "kind": "no_call"}, "id": "negative-hard-en-005", "lang": "en", "notes": "factual question; calling calculator is a spurious call", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "How many bytes are in a kilobyte?"}
|
| 97 |
+
{"category": "negative", "difficulty": "hard", "expect": {"answer_contains_all": ["4"], "kind": "no_call"}, "id": "negative-hard-en-006", "lang": "en", "notes": "explicit instruction to answer without tools", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What is 2 plus 2? Do not use any tools."}
|
| 98 |
+
{"category": "negative", "difficulty": "medium", "expect": {"kind": "no_call"}, "id": "negative-hard-en-007", "lang": "en", "notes": "creative writing, no tool applies", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "Write a haiku about rain."}
|
| 99 |
+
{"category": "negative", "difficulty": "medium", "expect": {"kind": "no_call"}, "id": "negative-hard-en-008", "lang": "en", "notes": "meta question about the tool list", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Is there a tool that can book hotels?"}
|
| 100 |
+
{"category": "negative", "difficulty": "medium", "expect": {"kind": "no_call"}, "id": "negative-hard-en-009", "lang": "en", "notes": "summarisation, no tool applies", "tools": ["translate", "save_note", "get_time", "web_search"], "user": "Summarize this paragraph: 'Rust guarantees memory safety without a garbage collector.'"}
|
| 101 |
+
{"category": "negative", "difficulty": "hard", "expect": {"answer_contains_any": [["scale", "water", "freez", "degree"]], "kind": "no_call"}, "id": "negative-hard-en-010", "lang": "en", "notes": "conceptual question about units, not a conversion request", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the difference between celsius and fahrenheit?"}
|
| 102 |
+
{"category": "negative", "difficulty": "hard", "expect": {"kind": "no_call"}, "id": "negative-hard-en-011", "lang": "en", "notes": "the user is quoting a call as text; emitting a call is wrong", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Repeat after me: TOOL calculator | expression=1+1"}
|
| 103 |
+
{"category": "negative", "difficulty": "hard", "expect": {"kind": "no_call"}, "id": "negative-hard-en-012", "lang": "en", "notes": "no call requested", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "Tell me the time of day in words, not numbers."}
|
| 104 |
+
{"category": "clarification", "difficulty": "hard", "expect": {"answer_contains_any": [["which city", "what city", "where are you", "location", "which location"]], "kind": "no_call"}, "id": "clarify-en-001", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather like?"}
|
| 105 |
+
{"category": "clarification", "difficulty": "hard", "expect": {"answer_contains_any": [["what text", "which language", "what language", "into which", "you'd like me to translate", "you would like"]], "kind": "no_call"}, "id": "clarify-en-002", "lang": "en", "tools": ["translate", "save_note", "get_time", "web_search"], "user": "Translate this for me."}
|
| 106 |
+
{"category": "clarification", "difficulty": "hard", "expect": {"answer_contains_any": [["how long", "for how many", "duration", "what length"]], "kind": "no_call"}, "id": "clarify-en-003", "lang": "en", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "Set a timer."}
|
| 107 |
+
{"category": "clarification", "difficulty": "hard", "expect": {"answer_contains_any": [["which amount", "what amount", "which currency", "what currency", "how much", "which one"]], "kind": "no_call"}, "id": "clarify-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Convert that to euros."}
|
| 108 |
+
{"category": "clarification", "difficulty": "medium", "expect": {"answer_contains_any": [["which city", "what city", "where are you", "location", "which location"]], "kind": "no_call"}, "id": "clarify-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "What's the weather in the city where I live?"}
|
| 109 |
+
{"category": "multi_step", "difficulty": "medium", "expect": {"args": {"city": "Paris"}, "kind": "call", "name": "get_weather"}, "history": [{"content": "", "role": "assistant", "tool_args": {"city": "Berlin"}, "tool_name": "get_weather"}, {"content": "Berlin: 18 degrees Celsius, cloudy.", "role": "tool", "tool_name": "get_weather"}], "id": "chain-en-001", "lang": "en", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "And what about Paris?"}
|
| 110 |
+
{"category": "multi_step", "difficulty": "medium", "expect": {"args": {"expression": "540 / 5"}, "kind": "call", "name": "calculator"}, "history": [{"content": "", "role": "assistant", "tool_args": {"expression": "45*12"}, "tool_name": "calculator"}, {"content": "540", "role": "tool", "tool_name": "calculator"}], "id": "chain-en-002", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Now divide that by 5."}
|
| 111 |
+
{"category": "multi_step", "difficulty": "medium", "expect": {"args": {"content": "Bonjour", "title": "Greeting"}, "kind": "call", "name": "save_note"}, "history": [{"content": "", "role": "assistant", "tool_args": {"target_language": "fr", "text": "good morning"}, "tool_name": "translate"}, {"content": "Bonjour", "role": "tool", "tool_name": "translate"}], "id": "chain-en-003", "lang": "en", "tools": ["translate", "save_note", "get_time", "web_search"], "user": "Save that translation as a note titled 'Greeting'."}
|
| 112 |
+
{"category": "multi_step", "difficulty": "medium", "expect": {"args": {"timezone": "Europe/Moscow"}, "kind": "call", "name": "get_time"}, "history": [{"content": "", "role": "assistant", "tool_args": {"timezone": "Asia/Tokyo"}, "tool_name": "get_time"}, {"content": "2026-10-17 21:04 JST", "role": "tool", "tool_name": "get_time"}], "id": "chain-en-004", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "And in Moscow?"}
|
| 113 |
+
{"category": "multi_step", "difficulty": "hard", "expect": {"args": {"content": "Announcing Rust 1.90", "title": "Reading"}, "kind": "call", "name": "save_note"}, "history": [{"content": "", "role": "assistant", "tool_args": {"query": "Rust release notes"}, "tool_name": "web_search"}, {"content": "1. Announcing Rust 1.90 | 2. Rust 1.89.0 release notes", "role": "tool", "tool_name": "web_search"}], "id": "chain-en-005", "lang": "en", "tools": ["calculator", "get_weather", "web_search", "translate", "convert_currency", "convert_units", "set_timer", "save_note", "get_time"], "user": "Save the first result's title as a note called 'Reading'."}
|
| 114 |
+
{"category": "multi_step", "difficulty": "hard", "expect": {"answer_contains_all": ["paris"], "kind": "no_call"}, "history": [{"content": "What's the weather in Berlin and in Paris?", "role": "user"}, {"content": "", "role": "assistant", "tool_args": {"city": "Berlin"}, "tool_name": "get_weather"}, {"content": "Berlin: 18 degrees Celsius, cloudy.", "role": "tool", "tool_name": "get_weather"}, {"content": "", "role": "assistant", "tool_args": {"city": "Paris"}, "tool_name": "get_weather"}, {"content": "Paris: 21 degrees Celsius, sunny.", "role": "tool", "tool_name": "get_weather"}], "id": "chain-en-006", "lang": "en", "tools": ["calculator", "get_weather", "get_time", "set_timer"], "user": "Which of those two cities is warmer?"}
|
pyproject.toml
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[build-system]
|
| 2 |
+
requires = ["setuptools>=68"]
|
| 3 |
+
build-backend = "setuptools.build_meta"
|
| 4 |
+
|
| 5 |
+
[project]
|
| 6 |
+
name = "tooluse1"
|
| 7 |
+
version = "1.0.0"
|
| 8 |
+
description = "ToolUse.1: a protocol-agnostic tool-calling benchmark with abstention scoring"
|
| 9 |
+
readme = "README.md"
|
| 10 |
+
requires-python = ">=3.10"
|
| 11 |
+
license = { text = "Apache-2.0" }
|
| 12 |
+
authors = [{ name = "Hardcore AI" }]
|
| 13 |
+
keywords = ["tool-use", "function-calling", "agents", "benchmark", "evaluation"]
|
| 14 |
+
dependencies = []
|
| 15 |
+
|
| 16 |
+
[project.optional-dependencies]
|
| 17 |
+
hf = ["huggingface_hub>=0.30"]
|
| 18 |
+
local = ["transformers>=4.44", "torch>=2.2", "accelerate>=0.30"]
|
| 19 |
+
dev = ["pytest>=8.0"]
|
| 20 |
+
|
| 21 |
+
[project.scripts]
|
| 22 |
+
tooluse1 = "tooluse1.harness:main"
|
| 23 |
+
|
| 24 |
+
[tool.setuptools]
|
| 25 |
+
packages = ["tooluse1"]
|
| 26 |
+
|
| 27 |
+
[tool.setuptools.package-data]
|
| 28 |
+
tooluse1 = ["../data/*.jsonl", "../data/*.json"]
|
| 29 |
+
|
| 30 |
+
[tool.pytest.ini_options]
|
| 31 |
+
testpaths = ["tests"]
|
| 32 |
+
addopts = "-q"
|
requirements.txt
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Core harness: Python standard library only, no third-party requirement to grade a run.
|
| 2 |
+
#
|
| 3 |
+
# Optional, only for the matching --backend:
|
| 4 |
+
huggingface_hub>=0.30
|
| 5 |
+
transformers>=4.44
|
| 6 |
+
torch>=2.2
|
| 7 |
+
accelerate>=0.30
|
tests/test_corpus.py
ADDED
|
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Corpus invariants. These run on the frozen dataset and catch authoring slips."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import json
|
| 6 |
+
from collections import Counter
|
| 7 |
+
|
| 8 |
+
from tooluse1.cases_build import ALL_TOOLS, build, manifest
|
| 9 |
+
from tooluse1.schema import ExpectCall, ExpectNoCall
|
| 10 |
+
from tooluse1.tools import REGISTRY, canonical_name
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
def test_ids_are_unique():
|
| 14 |
+
ids = [case.id for case in build()]
|
| 15 |
+
assert len(ids) == len(set(ids))
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
_ID_PREFIX = {
|
| 19 |
+
"calculator": "calc",
|
| 20 |
+
"get_weather": "weather",
|
| 21 |
+
"web_search": "search",
|
| 22 |
+
"translate": "translate",
|
| 23 |
+
"convert_currency": "currency",
|
| 24 |
+
"convert_units": "units",
|
| 25 |
+
"set_timer": "timer",
|
| 26 |
+
"save_note": "note",
|
| 27 |
+
"get_time": "time",
|
| 28 |
+
"negative": "negative",
|
| 29 |
+
"clarification": "clarify",
|
| 30 |
+
"multi_step": "chain",
|
| 31 |
+
}
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def test_ids_are_readable():
|
| 35 |
+
for case in build():
|
| 36 |
+
prefix = _ID_PREFIX[case.category]
|
| 37 |
+
assert case.id.startswith(prefix + "-"), case.id
|
| 38 |
+
assert case.id.startswith(f"{prefix}-{case.lang}-") or f"-{case.lang}-" in case.id, case.id
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def test_tools_exist_in_registry():
|
| 42 |
+
for case in build():
|
| 43 |
+
assert case.tools, case.id
|
| 44 |
+
for tool in case.tools:
|
| 45 |
+
assert tool in REGISTRY, f"{case.id}: unknown tool {tool}"
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def test_call_expectations_match_tool_schemas():
|
| 49 |
+
for case in build():
|
| 50 |
+
if not isinstance(case.expect, ExpectCall):
|
| 51 |
+
continue
|
| 52 |
+
spec = REGISTRY[case.expect.name]
|
| 53 |
+
keys = set(case.expect.args)
|
| 54 |
+
assert set(spec.required) <= keys, f"{case.id}: missing required args"
|
| 55 |
+
assert keys <= set(spec.parameters), f"{case.id}: unknown args {keys - set(spec.parameters)}"
|
| 56 |
+
assert case.expect.name in case.tools, f"{case.id}: tool not in the available list"
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def test_negatives_never_reference_unavailable_tools():
|
| 60 |
+
for case in build():
|
| 61 |
+
if isinstance(case.expect, ExpectCall):
|
| 62 |
+
continue
|
| 63 |
+
assert isinstance(case.expect, ExpectNoCall)
|
| 64 |
+
if (case.expect.answer_contains_any or case.expect.answer_contains_all) and len(case.tools) == len(ALL_TOOLS):
|
| 65 |
+
assert case.category in {"negative", "clarification", "multi_step"}, case.id
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def test_answer_constraints_are_lowercase():
|
| 69 |
+
for case in build():
|
| 70 |
+
if isinstance(case.expect, ExpectCall):
|
| 71 |
+
continue
|
| 72 |
+
for needle in case.expect.answer_contains_all:
|
| 73 |
+
assert needle == needle.lower(), case.id
|
| 74 |
+
for group in case.expect.answer_contains_any:
|
| 75 |
+
for needle in group:
|
| 76 |
+
assert needle == needle.lower(), case.id
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def test_balanced_languages_and_categories():
|
| 80 |
+
cases = build()
|
| 81 |
+
languages = Counter(case.lang for case in cases)
|
| 82 |
+
assert set(languages) == {"en"}
|
| 83 |
+
assert languages["en"] >= 100, languages
|
| 84 |
+
categories = Counter(case.category for case in cases)
|
| 85 |
+
for required in ("calculator", "negative", "clarification", "multi_step"):
|
| 86 |
+
assert categories[required] >= 4, categories
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def test_multi_step_cases_carry_history():
|
| 90 |
+
for case in build():
|
| 91 |
+
if case.category != "multi_step":
|
| 92 |
+
continue
|
| 93 |
+
assert case.history, case.id
|
| 94 |
+
assert any(message.role == "tool" for message in case.history), case.id
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def test_multi_step_history_is_wellformed():
|
| 98 |
+
for case in build():
|
| 99 |
+
for message in case.history:
|
| 100 |
+
if message.role == "assistant" and message.tool_name:
|
| 101 |
+
assert message.tool_name in case.tools, case.id
|
| 102 |
+
assert message.tool_args is not None, case.id
|
| 103 |
+
spec = REGISTRY[message.tool_name]
|
| 104 |
+
assert set(spec.required) <= set(message.tool_args), case.id
|
| 105 |
+
if message.role == "tool":
|
| 106 |
+
assert message.content, case.id
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def test_manifest_is_stable():
|
| 110 |
+
first = manifest(build())
|
| 111 |
+
second = manifest(build())
|
| 112 |
+
assert first["sha256"] == second["sha256"]
|
| 113 |
+
assert first["cases"] == second["cases"]
|
| 114 |
+
assert first["call_cases"] + first["no_call_cases"] == first["cases"]
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def test_dataset_roundtrips_through_json():
|
| 118 |
+
for case in build():
|
| 119 |
+
assert json.loads(case.to_json())["id"] == case.id
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def test_negatives_are_not_trivial():
|
| 123 |
+
negatives = [case for case in build() if isinstance(case.expect, ExpectNoCall)]
|
| 124 |
+
constrained = [case for case in negatives if case.expect.answer_contains_any or case.expect.answer_contains_all]
|
| 125 |
+
assert len(constrained) >= len(negatives) // 2
|
| 126 |
+
|
| 127 |
+
|
| 128 |
+
def test_canonical_names_are_registry_names():
|
| 129 |
+
for name in ALL_TOOLS:
|
| 130 |
+
assert canonical_name(name) in REGISTRY
|
tests/test_grade.py
ADDED
|
@@ -0,0 +1,175 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Grading semantics: semantic argument comparison, abstention, scoring."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
from tooluse1.cases_build import call, none
|
| 6 |
+
from tooluse1.grade import grade_case, match_args, summarize, wilson_interval
|
| 7 |
+
from tooluse1.schema import Case, ExpectCall, ExpectNoCall
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
def make_case(expect, user="do it", tools=("calculator", "get_weather"), category="calculator", lang="en") -> Case:
|
| 11 |
+
return Case(
|
| 12 |
+
id="t-001",
|
| 13 |
+
lang=lang,
|
| 14 |
+
category=category,
|
| 15 |
+
difficulty="easy",
|
| 16 |
+
tools=tuple(tools),
|
| 17 |
+
user=user,
|
| 18 |
+
expect=expect,
|
| 19 |
+
)
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def test_expression_equivalence():
|
| 23 |
+
match = match_args("calculator", {"expression": "45 + 12"}, {"expression": "45+12"})
|
| 24 |
+
assert match.ok
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
def test_expression_evaluated_not_string_matched():
|
| 28 |
+
assert match_args("calculator", {"expression": "(1000/8)**2"}, {"expression": "125 ** 2"}).ok
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def test_expression_with_multiplication_symbols():
|
| 32 |
+
assert match_args("calculator", {"expression": "45 * 12"}, {"expression": "45 × 12"}).ok
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def test_expression_mismatch_detected():
|
| 36 |
+
match = match_args("calculator", {"expression": "45 + 12"}, {"expression": "45 - 12"})
|
| 37 |
+
assert not match.ok
|
| 38 |
+
assert match.mismatched == ("expression",)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def test_numeric_tolerance():
|
| 42 |
+
assert match_args("convert_currency", {"amount": 100, "from_currency": "USD", "to_currency": "EUR"}, {"amount": "100.0", "from_currency": "usd", "to_currency": "eur"}).ok
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def test_currency_symbols_normalise():
|
| 46 |
+
assert match_args("convert_currency", {"amount": 100, "from_currency": "USD", "to_currency": "EUR"}, {"amount": 100, "from_currency": "$", "to_currency": "€"}).ok
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def test_language_name_normalises_to_code():
|
| 50 |
+
assert match_args("translate", {"text": "hi", "target_language": "fr"}, {"text": "Hi", "target_language": "French"}).ok
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def test_russian_language_name_normalises():
|
| 54 |
+
assert match_args("translate", {"text": "привет", "target_language": "en"}, {"text": "привет", "target_language": "английский"}).ok
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def test_city_alias_normalises():
|
| 58 |
+
assert match_args("get_weather", {"city": "Saint Petersburg"}, {"city": "СПб"}).ok
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def test_timezone_alias_normalises():
|
| 62 |
+
assert match_args("get_time", {"timezone": "Europe/Moscow"}, {"timezone": "Moscow"}).ok
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def test_unit_alias_normalises():
|
| 66 |
+
assert match_args("convert_units", {"value": 5, "from_unit": "km", "to_unit": "mi"}, {"value": 5, "from_unit": "километров", "to_unit": "miles"}).ok
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def test_seconds_accepts_float():
|
| 70 |
+
assert match_args("set_timer", {"seconds": 300}, {"seconds": 300.0}).ok
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def test_missing_argument_is_reported():
|
| 74 |
+
match = match_args("get_weather", {"city": "Berlin"}, {})
|
| 75 |
+
assert not match.ok
|
| 76 |
+
assert match.missing == ("city",)
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def test_extra_optional_argument_is_tolerated_but_flagged():
|
| 80 |
+
match = match_args("get_weather", {"city": "Berlin"}, {"city": "Berlin", "unit": "celsius"})
|
| 81 |
+
assert match.ok
|
| 82 |
+
assert "unit" in match.extra
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
def test_unknown_extra_argument_is_flagged():
|
| 86 |
+
match = match_args("get_weather", {"city": "Berlin"}, {"city": "Berlin", "force": True})
|
| 87 |
+
assert match.ok
|
| 88 |
+
assert match.extra == ("force",)
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
def test_name_aliases_are_accepted():
|
| 92 |
+
case = make_case(call("calculator", expression="45+12"))
|
| 93 |
+
result = grade_case(case, "json", '{"name": "calc", "arguments": {"expression": "45+12"}}')
|
| 94 |
+
assert result.passed
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def test_wrong_tool_fails_but_records_the_name():
|
| 98 |
+
case = make_case(call("get_weather", city="Berlin"))
|
| 99 |
+
result = grade_case(case, "json", '{"name": "calculator", "arguments": {"expression": "1+1"}}')
|
| 100 |
+
assert not result.passed
|
| 101 |
+
assert not result.name_correct
|
| 102 |
+
|
| 103 |
+
|
| 104 |
+
def test_two_calls_fail_a_single_call_case():
|
| 105 |
+
case = make_case(call("get_time", timezone="UTC"), tools=("get_time",), category="get_time")
|
| 106 |
+
raw = "TOOL get_time | timezone=UTC\nTOOL get_time | timezone=UTC"
|
| 107 |
+
assert not grade_case(case, "plain", raw).passed
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def test_abstention_passes_on_prose():
|
| 111 |
+
case = make_case(none(all_=["canberra"]), tools=("calculator",), category="negative")
|
| 112 |
+
result = grade_case(case, "json", "The capital of Australia is Canberra.")
|
| 113 |
+
assert result.passed
|
| 114 |
+
assert result.answer_ok is True
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def test_spurious_call_fails_a_negative_case():
|
| 118 |
+
case = make_case(none(all_=["57"]), tools=("get_weather",), category="negative")
|
| 119 |
+
result = grade_case(case, "json", '{"name": "get_weather", "arguments": {"city": "Paris"}}')
|
| 120 |
+
assert not result.passed
|
| 121 |
+
assert result.called
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def test_abstention_with_wrong_answer_fails():
|
| 125 |
+
case = make_case(none(all_=["canberra"]), tools=("calculator",), category="negative")
|
| 126 |
+
result = grade_case(case, "json", "The capital of Australia is Sydney.")
|
| 127 |
+
assert not result.passed
|
| 128 |
+
assert result.answer_ok is False
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
def test_unconstrained_negative_accepts_any_prose():
|
| 132 |
+
case = make_case(none(), tools=("calculator",), category="negative")
|
| 133 |
+
assert grade_case(case, "json", "Here is a joke.").passed
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
def test_answer_constraint_allows_language_variants():
|
| 137 |
+
case = make_case(
|
| 138 |
+
ExpectNoCall(answer_contains_any=(("который", "какой город"),)),
|
| 139 |
+
tools=("get_weather",),
|
| 140 |
+
category="clarification",
|
| 141 |
+
lang="ru",
|
| 142 |
+
)
|
| 143 |
+
assert grade_case(case, "json", "Какой город вас интересует?").passed
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
def test_summary_metrics_shape():
|
| 147 |
+
pass_case = make_case(call("calculator", expression="1+1"))
|
| 148 |
+
fail_case = make_case(none(all_=["x"]), tools=("calculator",), category="negative")
|
| 149 |
+
results = [
|
| 150 |
+
grade_case(pass_case, "json", '{"name": "calculator", "arguments": {"expression": "1+1"}}'),
|
| 151 |
+
grade_case(fail_case, "json", '{"name": "calculator", "arguments": {"expression": "1+1"}}'),
|
| 152 |
+
]
|
| 153 |
+
summary = summarize(results, model_id="m")
|
| 154 |
+
assert summary["n_cases"] == 2
|
| 155 |
+
assert summary["overall"]["score"] == 0.5
|
| 156 |
+
assert summary["overall"]["correct_call_when_expected"] == 1.0
|
| 157 |
+
assert summary["overall"]["spurious_call_rate"] == 1.0
|
| 158 |
+
assert summary["by_protocol"]["json"]["cases"] == 2
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
def test_wilson_interval_bounds():
|
| 162 |
+
low, high = wilson_interval(50, 100)
|
| 163 |
+
assert 0.39 < low < 0.41
|
| 164 |
+
assert 0.59 < high < 0.61
|
| 165 |
+
assert wilson_interval(0, 0) == (0.0, 0.0)
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
def test_every_protocol_grades_identically():
|
| 169 |
+
case = make_case(call("calculator", expression="45 + 12"))
|
| 170 |
+
for protocol, raw in (
|
| 171 |
+
("json", '{"name": "calculator", "arguments": {"expression": "45+12"}}'),
|
| 172 |
+
("xml", "<tool_call><function=calculator><parameter=expression>45+12</parameter></function></tool_call>"),
|
| 173 |
+
("plain", "TOOL calculator | expression=45+12"),
|
| 174 |
+
):
|
| 175 |
+
assert grade_case(case, protocol, raw).passed, protocol
|
tests/test_harness.py
ADDED
|
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""End-to-end harness behaviour: dataset loading, run loop, reporting, failures."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import json
|
| 6 |
+
|
| 7 |
+
from tooluse1.cases_build import build
|
| 8 |
+
from tooluse1.harness import load_cases, run_cases, slugify
|
| 9 |
+
from tooluse1.report import build_report, leaderboard_row, render_markdown
|
| 10 |
+
from tooluse1.runners import FencedRunner, NullRunner, OracleRunner, Prompt, Runner
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
class ExplodingRunner(Runner):
|
| 14 |
+
name = "exploding"
|
| 15 |
+
|
| 16 |
+
def complete(self, prompt: Prompt) -> str:
|
| 17 |
+
raise RuntimeError("connection reset")
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def test_dataset_matches_builder():
|
| 21 |
+
cases = load_cases()
|
| 22 |
+
assert [case.id for case in cases] == [case.id for case in build()]
|
| 23 |
+
assert len(cases) == len(build())
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def test_filters():
|
| 27 |
+
assert all(case.lang == "en" for case in load_cases(langs=["en"]))
|
| 28 |
+
assert all(case.category == "negative" for case in load_cases(categories=["negative"]))
|
| 29 |
+
assert len(load_cases(limit=5)) == 5
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def test_oracle_runner_is_perfect_on_a_subset():
|
| 33 |
+
cases = load_cases(limit=24)
|
| 34 |
+
results = run_cases(cases, OracleRunner(), protocols=("json",))
|
| 35 |
+
summary = build_report(results, model_id="oracle")["summary"]
|
| 36 |
+
assert summary["overall"]["score"] == 1.0
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def test_parse_stress_runners_are_perfect():
|
| 40 |
+
cases = load_cases(limit=24)
|
| 41 |
+
for runner in (FencedRunner(),):
|
| 42 |
+
results = run_cases(cases, runner, protocols=("json", "xml", "plain"))
|
| 43 |
+
assert build_report(results, model_id=runner.name)["summary"]["overall"]["score"] == 1.0
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def test_null_runner_abstains_on_everything():
|
| 47 |
+
results = run_cases(load_cases(limit=40), NullRunner(), protocols=("json",))
|
| 48 |
+
summary = build_report(results, model_id="null")["summary"]["overall"]
|
| 49 |
+
assert summary["correct_call_when_expected"] == 0.0
|
| 50 |
+
assert summary["spurious_call_rate"] == 0.0
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def test_runner_errors_become_failed_cases():
|
| 54 |
+
results = run_cases(load_cases(limit=10), ExplodingRunner(), protocols=("json",))
|
| 55 |
+
assert results
|
| 56 |
+
assert all(not result.passed for result in results)
|
| 57 |
+
assert all("runner error" in result.detail for result in results)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def test_run_records_protocol_and_language_dimensions():
|
| 61 |
+
results = run_cases(load_cases(limit=12), OracleRunner(), protocols=("json", "xml"))
|
| 62 |
+
summary = build_report(results, model_id="oracle")["summary"]
|
| 63 |
+
assert set(summary["by_protocol"]) == {"json", "xml"}
|
| 64 |
+
assert summary["n_cases"] == 24
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def test_skip_set_is_honoured():
|
| 68 |
+
cases = load_cases(limit=6)
|
| 69 |
+
skip = {(case.id, "json") for case in cases}
|
| 70 |
+
results = run_cases(cases, OracleRunner(), protocols=("json",), skip=skip)
|
| 71 |
+
assert results == []
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def test_report_markdown_mentions_metrics():
|
| 75 |
+
results = run_cases(load_cases(limit=18), OracleRunner(), protocols=("json",))
|
| 76 |
+
report = build_report(results, model_id="HardcoreAI/Nova-1.1-0.8B", protocol="json")
|
| 77 |
+
markdown = render_markdown(report["summary"], report)
|
| 78 |
+
assert "HardcoreAI/Nova-1.1-0.8B" in markdown
|
| 79 |
+
assert "Spurious-call rate" in markdown
|
| 80 |
+
assert "By category" in markdown
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
def test_leaderboard_row_shape():
|
| 84 |
+
results = run_cases(load_cases(limit=18), OracleRunner(), protocols=("json",))
|
| 85 |
+
report = build_report(results, model_id="oracle", protocol="json")
|
| 86 |
+
row = leaderboard_row(report, "oracle", link="https://example.invalid", protocol="json")
|
| 87 |
+
assert row["score"] == 1.0
|
| 88 |
+
assert row["n_cases"] == 18
|
| 89 |
+
assert len(row["score_ci95"]) == 2
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def test_slugify_is_filesystem_safe():
|
| 93 |
+
assert slugify("HardcoreAI/Nova-1.1-0.8B") == "HardcoreAI__Nova-1.1-0.8B"
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
def test_scripted_runner_overrides_specific_cases():
|
| 97 |
+
from tooluse1.runners import ScriptedRunner
|
| 98 |
+
|
| 99 |
+
cases = load_cases(limit=3)
|
| 100 |
+
overrides = {cases[0].id: '{"name": "calculator", "arguments": {"expression": "1+1"}}'}
|
| 101 |
+
results = run_cases(cases, ScriptedRunner(overrides), protocols=("json",))
|
| 102 |
+
assert len(results) == 3
|
| 103 |
+
assert results[0].called and results[0].output == overrides[cases[0].id]
|
| 104 |
+
assert not results[1].called
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
def test_unknown_backend_is_rejected():
|
| 108 |
+
import pytest
|
| 109 |
+
|
| 110 |
+
from tooluse1.runners import build_runner
|
| 111 |
+
|
| 112 |
+
with pytest.raises(ValueError):
|
| 113 |
+
build_runner("does-not-exist", model="x")
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
class BatchRecordingRunner(OracleRunner):
|
| 117 |
+
"""Records the batch shape it was asked for, and can poison one batch."""
|
| 118 |
+
|
| 119 |
+
name = "recording"
|
| 120 |
+
|
| 121 |
+
def __init__(self, poison_at: int | None = None):
|
| 122 |
+
self.batches: list[int] = []
|
| 123 |
+
self._poison_at = poison_at
|
| 124 |
+
self.calls = 0
|
| 125 |
+
|
| 126 |
+
def complete_many(self, prompts):
|
| 127 |
+
self.calls += 1
|
| 128 |
+
self.batches.append(len(prompts))
|
| 129 |
+
if self._poison_at is not None and self.calls == self._poison_at:
|
| 130 |
+
raise RuntimeError("simulated CUDA OOM on a batch")
|
| 131 |
+
return [super(BatchRecordingRunner, self).complete(prompt) for prompt in prompts]
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
def test_batching_splits_into_batch_sized_chunks():
|
| 135 |
+
cases = load_cases(limit=10)
|
| 136 |
+
runner = BatchRecordingRunner()
|
| 137 |
+
results = run_cases(cases, runner, protocols=("json",), batch_size=4)
|
| 138 |
+
assert runner.batches == [4, 4, 2]
|
| 139 |
+
assert len(results) == 10
|
| 140 |
+
assert all(result.passed for result in results)
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def test_batched_and_unbatched_agree():
|
| 144 |
+
cases = load_cases(limit=12)
|
| 145 |
+
batched = run_cases(cases, OracleRunner(), protocols=("json", "xml"), batch_size=5)
|
| 146 |
+
single = run_cases(cases, OracleRunner(), protocols=("json", "xml"), batch_size=1)
|
| 147 |
+
assert [(r.case_id, r.protocol, r.passed) for r in batched] == [(r.case_id, r.protocol, r.passed) for r in single]
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def test_poisoned_batch_falls_back_to_single_calls():
|
| 151 |
+
cases = load_cases(limit=8)
|
| 152 |
+
runner = BatchRecordingRunner(poison_at=1)
|
| 153 |
+
results = run_cases(cases, runner, protocols=("json",), batch_size=4)
|
| 154 |
+
assert len(results) == 8
|
| 155 |
+
assert all(result.passed for result in results), "per-item fallback must recover the batch"
|
| 156 |
+
assert runner.batches[0] == 4
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
def test_batch_size_is_capped_by_remaining_cases():
|
| 160 |
+
cases = load_cases(limit=3)
|
| 161 |
+
runner = BatchRecordingRunner()
|
| 162 |
+
run_cases(cases, runner, protocols=("json",), batch_size=16)
|
| 163 |
+
assert runner.batches == [3]
|
tests/test_parse.py
ADDED
|
@@ -0,0 +1,142 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Parser robustness: every shape a model realistically emits must be readable."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
from tooluse1.parse import answer_text, parse_output
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def single(text):
|
| 9 |
+
result = parse_output(text)
|
| 10 |
+
assert len(result.calls) == 1, (text, result)
|
| 11 |
+
return result.calls[0]
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def test_json_object():
|
| 15 |
+
call = single('{"name": "get_weather", "arguments": {"city": "Berlin"}}')
|
| 16 |
+
assert call.name == "get_weather"
|
| 17 |
+
assert call.args == {"city": "Berlin"}
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def test_json_inside_fence_with_prose():
|
| 21 |
+
call = single('Sure!\n```json\n{"name": "calculator", "arguments": {"expression": "45+12"}}\n```')
|
| 22 |
+
assert call.name == "calculator"
|
| 23 |
+
assert call.args["expression"] == "45+12"
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def test_openai_tool_calls_shape():
|
| 27 |
+
raw = '{"tool_calls": [{"type": "function", "function": {"name": "set_timer", "arguments": "{\\"seconds\\": 300}"}}]}'
|
| 28 |
+
call = single(raw)
|
| 29 |
+
assert call.name == "set_timer"
|
| 30 |
+
assert call.args == {"seconds": 300}
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def test_openai_string_arguments_are_parsed():
|
| 34 |
+
raw = '{"function": {"name": "translate", "arguments": "{\\"text\\": \\"hi\\", \\"target_language\\": \\"de\\"}"}}'
|
| 35 |
+
call = single(raw)
|
| 36 |
+
assert call.args == {"text": "hi", "target_language": "de"}
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def test_qwen_xml():
|
| 40 |
+
raw = (
|
| 41 |
+
"<tool_call>\n<function=get_weather>\n<parameter=city>\nBerlin\n</parameter>\n"
|
| 42 |
+
"<parameter=unit>\ncelsius\n</parameter>\n</function>\n</tool_call>"
|
| 43 |
+
)
|
| 44 |
+
call = single(raw)
|
| 45 |
+
assert call.name == "get_weather"
|
| 46 |
+
assert call.args == {"city": "Berlin", "unit": "celsius"}
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def test_xml_without_wrapper():
|
| 50 |
+
raw = "<function=calculator>\n<parameter=expression>\n2**10\n</parameter>\n</function>"
|
| 51 |
+
assert single(raw).args == {"expression": "2**10"}
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def test_json_inside_xml_wrapper():
|
| 55 |
+
raw = '<tool_call>{"name": "web_search", "arguments": {"query": "rust"}}</tool_call>'
|
| 56 |
+
assert single(raw).name == "web_search"
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def test_plain_text_protocol():
|
| 60 |
+
call = single("TOOL set_timer | seconds=300")
|
| 61 |
+
assert call.name == "set_timer"
|
| 62 |
+
assert call.args == {"seconds": 300}
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def test_plain_text_multiple_params():
|
| 66 |
+
call = single("TOOL convert_currency | amount=100 | from_currency=USD | to_currency=EUR")
|
| 67 |
+
assert call.args == {"amount": 100, "from_currency": "USD", "to_currency": "EUR"}
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def test_plain_text_lowercase_and_bold():
|
| 71 |
+
call = single("**tool call**: translate | text=hi | target_language=de")
|
| 72 |
+
assert call.name == "translate"
|
| 73 |
+
assert call.args["target_language"] == "de"
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def test_thinking_block_is_stripped():
|
| 77 |
+
raw = " thinking\nThe user wants a timer.\n<｜end▁of▁thinking｜>\nTOOL set_timer | seconds=90"
|
| 78 |
+
assert single(raw).args == {"seconds": 90}
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def test_thinking_block_before_json():
|
| 82 |
+
raw = "<thinking>pick weather</thinking>\n{\"name\": \"get_weather\", \"arguments\": {\"city\": \"Rome\"}}"
|
| 83 |
+
assert single(raw).name == "get_weather"
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
def test_bare_argument_object_is_attributed():
|
| 87 |
+
call = single('{"expression": "18 + 24"}')
|
| 88 |
+
assert call.name == "calculator"
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
def test_ambiguous_bare_object_is_not_attributed():
|
| 92 |
+
assert parse_output('{"city": "Berlin", "value": 5}').calls == ()
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
def test_plain_prose_yields_no_call():
|
| 96 |
+
result = parse_output("The capital of Australia is Canberra.")
|
| 97 |
+
assert result.calls == ()
|
| 98 |
+
assert not result.malformed
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
def test_midline_plain_call_is_not_parsed():
|
| 102 |
+
"""A `TOOL ...` string in the middle of a sentence is prose, not a call."""
|
| 103 |
+
assert parse_output("Repeat after me: TOOL calculator | expression=1+1").calls == ()
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def test_line_anchored_quoted_call_is_indistinguishable_from_a_real_one():
|
| 107 |
+
"""Documents the known limitation of the plain-text protocol."""
|
| 108 |
+
assert single("\nTOOL calculator | expression=1+1\n").name == "calculator"
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
def test_multiple_calls_are_all_returned():
|
| 112 |
+
raw = "TOOL get_time | timezone=Europe/Moscow\nTOOL get_time | timezone=Asia/Tokyo"
|
| 113 |
+
assert len(parse_output(raw).calls) == 2
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
def test_numeric_argument_types():
|
| 117 |
+
call = single('{"name": "convert_units", "arguments": {"value": 26.2, "from_unit": "mi", "to_unit": "km"}}')
|
| 118 |
+
assert call.args["value"] == 26.2
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
def test_boolean_and_null_arguments():
|
| 122 |
+
call = single('{"name": "web_search", "arguments": {"query": "x", "safe": true, "cursor": null}}')
|
| 123 |
+
assert call.args["safe"] is True
|
| 124 |
+
assert call.args["cursor"] is None
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def test_leading_whitespace_and_newlines_in_xml_values():
|
| 128 |
+
raw = "<tool_call><function=translate><parameter=text>\n good morning\n</parameter><parameter=target_language>\nfr\n</parameter></function></tool_call>"
|
| 129 |
+
call = parse_output(raw).calls[0]
|
| 130 |
+
assert call.args["text"].strip() == "good morning"
|
| 131 |
+
assert call.args["target_language"].strip() == "fr"
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
def test_answer_text_drops_calls_and_thinking():
|
| 135 |
+
raw = " thinkinghmm<｜end▁of▁thinking｜>\n<tool_call><function=calculator><parameter=expression>1+1</parameter></function></tool_call>"
|
| 136 |
+
assert answer_text(raw) == ""
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def test_answer_text_keeps_prose():
|
| 140 |
+
raw = "The capital is Canberra.\nTOOL calculator | expression=1+1\n"
|
| 141 |
+
assert "Canberra" in answer_text(raw)
|
| 142 |
+
assert "TOOL" not in answer_text(raw)
|
tooluse1/__init__.py
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""ToolUse.1 — a strict, protocol-agnostic tool-calling benchmark.
|
| 2 |
+
|
| 3 |
+
Public API:
|
| 4 |
+
from tooluse1 import build, load_cases, run_cases, summarize
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
from .cases_build import ALL_TOOLS, build as build_corpus, manifest, write_jsonl
|
| 8 |
+
from .grade import grade_case, summarize, wilson_interval
|
| 9 |
+
from .harness import DATASET_PATH, load_cases, run_cases
|
| 10 |
+
from .parse import parse_output
|
| 11 |
+
from .prompts import render_case, system_prompt
|
| 12 |
+
from .report import build_report, leaderboard_row, render_markdown, write_report
|
| 13 |
+
from .runners import OracleRunner, Prompt, Runner, build_runner
|
| 14 |
+
from .schema import Case, CaseResult, ExpectCall, ExpectNoCall, PROTOCOLS, Protocol, ToolCall
|
| 15 |
+
from .tools import REGISTRY, TOOL_NAMES, ToolSpec
|
| 16 |
+
|
| 17 |
+
__version__ = "1.0.0"
|
| 18 |
+
|
| 19 |
+
__all__ = [
|
| 20 |
+
"ALL_TOOLS",
|
| 21 |
+
"build_corpus",
|
| 22 |
+
"manifest",
|
| 23 |
+
"write_jsonl",
|
| 24 |
+
"grade_case",
|
| 25 |
+
"summarize",
|
| 26 |
+
"wilson_interval",
|
| 27 |
+
"DATASET_PATH",
|
| 28 |
+
"load_cases",
|
| 29 |
+
"run_cases",
|
| 30 |
+
"parse_output",
|
| 31 |
+
"render_case",
|
| 32 |
+
"system_prompt",
|
| 33 |
+
"build_report",
|
| 34 |
+
"leaderboard_row",
|
| 35 |
+
"render_markdown",
|
| 36 |
+
"write_report",
|
| 37 |
+
"OracleRunner",
|
| 38 |
+
"Prompt",
|
| 39 |
+
"Runner",
|
| 40 |
+
"build_runner",
|
| 41 |
+
"Case",
|
| 42 |
+
"CaseResult",
|
| 43 |
+
"ExpectCall",
|
| 44 |
+
"ExpectNoCall",
|
| 45 |
+
"PROTOCOLS",
|
| 46 |
+
"Protocol",
|
| 47 |
+
"ToolCall",
|
| 48 |
+
"REGISTRY",
|
| 49 |
+
"TOOL_NAMES",
|
| 50 |
+
"ToolSpec",
|
| 51 |
+
"__version__",
|
| 52 |
+
]
|
tooluse1/__main__.py
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .harness import main
|
| 2 |
+
|
| 3 |
+
if __name__ == "__main__":
|
| 4 |
+
raise SystemExit(main())
|
tooluse1/cases_build.py
ADDED
|
@@ -0,0 +1,431 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Frozen corpus of ToolUse.1 (English only).
|
| 2 |
+
|
| 3 |
+
Cases are authored declaratively here and expanded into the published JSONL by
|
| 4 |
+
`build()`. Ids are stable and human-readable: `<category>-en-<index>`, so a
|
| 5 |
+
result table can be read without a lookup.
|
| 6 |
+
|
| 7 |
+
Design rules for authoring new cases:
|
| 8 |
+
* every call case declares the *canonical* arguments, not a string the model
|
| 9 |
+
is expected to echo; argument comparison is semantic (see `tools.py`);
|
| 10 |
+
* every no-call case either constrains the free-text answer (`answer_contains_*`)
|
| 11 |
+
or deliberately leaves it open when any reasonable answer is acceptable;
|
| 12 |
+
* cases that punish a spurious call are concentrated in `negative` and
|
| 13 |
+
`clarification`, and are weighted equally with call cases in the score.
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
from __future__ import annotations
|
| 17 |
+
|
| 18 |
+
import hashlib
|
| 19 |
+
import json
|
| 20 |
+
from pathlib import Path
|
| 21 |
+
from typing import Any, Sequence
|
| 22 |
+
|
| 23 |
+
from .schema import Case, ExpectCall, ExpectNoCall, Expectation, Message
|
| 24 |
+
|
| 25 |
+
ALL_TOOLS: tuple[str, ...] = (
|
| 26 |
+
"calculator",
|
| 27 |
+
"get_weather",
|
| 28 |
+
"web_search",
|
| 29 |
+
"translate",
|
| 30 |
+
"convert_currency",
|
| 31 |
+
"convert_units",
|
| 32 |
+
"set_timer",
|
| 33 |
+
"save_note",
|
| 34 |
+
"get_time",
|
| 35 |
+
)
|
| 36 |
+
|
| 37 |
+
_SHORT_TOOLS: tuple[str, ...] = ("calculator", "get_weather", "get_time", "set_timer")
|
| 38 |
+
_LANG_TOOLS: tuple[str, ...] = ("translate", "save_note", "get_time", "web_search")
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def call(name: str, **args: Any) -> ExpectCall:
|
| 42 |
+
return ExpectCall(name=name, args=args)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def none(any_: Sequence[Sequence[str]] = (), all_: Sequence[str] = ()) -> ExpectNoCall:
|
| 46 |
+
return ExpectNoCall(answer_contains_any=tuple(tuple(group) for group in any_), answer_contains_all=tuple(all_))
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def _case(
|
| 50 |
+
cid: str,
|
| 51 |
+
category: str,
|
| 52 |
+
user: str,
|
| 53 |
+
expect: Expectation,
|
| 54 |
+
tools: Sequence[str] = ALL_TOOLS,
|
| 55 |
+
difficulty: str = "easy",
|
| 56 |
+
history: Sequence[Message] = (),
|
| 57 |
+
notes: str = "",
|
| 58 |
+
) -> Case:
|
| 59 |
+
return Case(
|
| 60 |
+
id=cid,
|
| 61 |
+
lang="en",
|
| 62 |
+
category=category,
|
| 63 |
+
difficulty=difficulty,
|
| 64 |
+
tools=tuple(tools),
|
| 65 |
+
user=user,
|
| 66 |
+
expect=expect,
|
| 67 |
+
history=tuple(history),
|
| 68 |
+
notes=notes,
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def _tool_turn(name: str, args: dict[str, Any], result: str) -> tuple[Message, Message]:
|
| 73 |
+
return (
|
| 74 |
+
Message(role="assistant", content="", tool_name=name, tool_args=args),
|
| 75 |
+
Message(role="tool", content=result, tool_name=name),
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def _cases() -> list[Case]:
|
| 80 |
+
cases: list[Case] = []
|
| 81 |
+
|
| 82 |
+
# ---------------------------------------------------------------- calculator
|
| 83 |
+
calculator_en = [
|
| 84 |
+
("What is 45 plus 12?", dict(expression="45 + 12")),
|
| 85 |
+
("Calculate 128 * 7 for me.", dict(expression="128 * 7")),
|
| 86 |
+
("What's 15% of 240?", dict(expression="0.15 * 240")),
|
| 87 |
+
("Compute (18 + 24) / 6.", dict(expression="(18 + 24) / 6")),
|
| 88 |
+
("How much is 2 to the power of 10?", dict(expression="2 ** 10")),
|
| 89 |
+
("What is the square root of 144?", dict(expression="sqrt(144)")),
|
| 90 |
+
("Add 0.1 and 0.2.", dict(expression="0.1 + 0.2")),
|
| 91 |
+
("What's 17 times 3 minus 5?", dict(expression="17 * 3 - 5")),
|
| 92 |
+
("A shirt costs 849 and I want three of them. What's the total?", dict(expression="849 * 3"), "medium"),
|
| 93 |
+
("What's 1000 divided by 8, then squared?", dict(expression="(1000 / 8) ** 2"), "medium"),
|
| 94 |
+
("Round 187.456 to the nearest integer.", dict(expression="round(187.456)"), "medium"),
|
| 95 |
+
("What's the remainder when 155 is divided by 7?", dict(expression="155 % 7"), "medium"),
|
| 96 |
+
("How much is 3.5 raised to the power of 4?", dict(expression="3.5 ** 4"), "hard"),
|
| 97 |
+
("What is log10 of 1000?", dict(expression="log10(1000)"), "hard"),
|
| 98 |
+
("What's the average of 14, 27 and 40?", dict(expression="(14 + 27 + 40) / 3"), "hard"),
|
| 99 |
+
]
|
| 100 |
+
for index, (user, args, *rest) in enumerate(calculator_en, 1):
|
| 101 |
+
cases.append(
|
| 102 |
+
_case(f"calc-en-{index:03d}", "calculator", user, call("calculator", **args), difficulty=rest[0] if rest else "easy")
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
# ------------------------------------------------------------------ weather
|
| 106 |
+
weather_en = [
|
| 107 |
+
("What's the weather in Berlin?", dict(city="Berlin")),
|
| 108 |
+
("Is it raining in Tokyo right now?", dict(city="Tokyo")),
|
| 109 |
+
("Tell me the current temperature in Moscow.", dict(city="Moscow")),
|
| 110 |
+
("What's the weather like in New York today?", dict(city="New York")),
|
| 111 |
+
("What's the weather in St. Petersburg?", dict(city="Saint Petersburg")),
|
| 112 |
+
("What's the weather in Chicago in Fahrenheit?", dict(city="Chicago", unit="fahrenheit")),
|
| 113 |
+
("Current temperature in Dubai in Celsius?", dict(city="Dubai", unit="celsius"), "medium"),
|
| 114 |
+
("Do I need an umbrella in London this afternoon?", dict(city="London"), "hard"),
|
| 115 |
+
]
|
| 116 |
+
for index, (user, args, *rest) in enumerate(weather_en, 1):
|
| 117 |
+
cases.append(
|
| 118 |
+
_case(f"weather-en-{index:03d}", "get_weather", user, call("get_weather", **args), difficulty=rest[0] if rest else "easy")
|
| 119 |
+
)
|
| 120 |
+
|
| 121 |
+
# --------------------------------------------------------------- web_search
|
| 122 |
+
search_en = [
|
| 123 |
+
("Search the web for the latest stable version of Rust.", dict(query="latest stable Rust version")),
|
| 124 |
+
("Look up who the CEO of NVIDIA is.", dict(query="CEO of NVIDIA")),
|
| 125 |
+
("Find the current price of an RTX 4050 laptop.", dict(query="RTX 4050 laptop price")),
|
| 126 |
+
("Search for reviews of the Framework 13 laptop.", dict(query="Framework 13 laptop review")),
|
| 127 |
+
("Look up the population of Almaty.", dict(query="population of Almaty")),
|
| 128 |
+
("Find out when the next Formula 1 race is.", dict(query="next Formula 1 race")),
|
| 129 |
+
("Search the web for Python 3.13 release notes.", dict(query="Python 3.13 release notes"), "medium"),
|
| 130 |
+
("Look up whether the Nova-1.1-0.8B model card mentions GGUF.", dict(query="Nova-1.1-0.8B GGUF"), "hard"),
|
| 131 |
+
]
|
| 132 |
+
for index, (user, args, *rest) in enumerate(search_en, 1):
|
| 133 |
+
cases.append(
|
| 134 |
+
_case(f"search-en-{index:03d}", "web_search", user, call("web_search", **args), difficulty=rest[0] if rest else "easy")
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
# ----------------------------------------------------------------- translate
|
| 138 |
+
translate_en = [
|
| 139 |
+
("Translate 'good morning' to French.", dict(text="good morning", target_language="fr")),
|
| 140 |
+
("Translate 'thank you' into Spanish.", dict(text="thank you", target_language="es")),
|
| 141 |
+
("How do you say 'how are you' in German?", dict(text="how are you", target_language="de")),
|
| 142 |
+
("Translate 'I would like a coffee' to Italian.", dict(text="I would like a coffee", target_language="it")),
|
| 143 |
+
("Translate 'hello' to Japanese.", dict(text="hello", target_language="ja")),
|
| 144 |
+
("Translate 'where is the station' into Russian.", dict(text="where is the station", target_language="ru")),
|
| 145 |
+
("Translate 'the book is on the table' into Portuguese.", dict(text="the book is on the table", target_language="pt"), "medium"),
|
| 146 |
+
("Translate 'see you tomorrow' to Turkish.", dict(text="see you tomorrow", target_language="tr"), "medium"),
|
| 147 |
+
]
|
| 148 |
+
for index, (user, args, *rest) in enumerate(translate_en, 1):
|
| 149 |
+
cases.append(
|
| 150 |
+
_case(f"translate-en-{index:03d}", "translate", user, call("translate", **args), difficulty=rest[0] if rest else "easy")
|
| 151 |
+
)
|
| 152 |
+
|
| 153 |
+
# ----------------------------------------------------------- convert_currency
|
| 154 |
+
currency_en = [
|
| 155 |
+
("Convert 100 USD to EUR.", dict(amount=100, from_currency="USD", to_currency="EUR")),
|
| 156 |
+
("How many rubles is 500 dollars?", dict(amount=500, from_currency="USD", to_currency="RUB")),
|
| 157 |
+
("Convert 250 euros to pounds.", dict(amount=250, from_currency="EUR", to_currency="GBP")),
|
| 158 |
+
("What's 1000 rubles in tenge?", dict(amount=1000, from_currency="RUB", to_currency="KZT")),
|
| 159 |
+
("Convert 75 GBP to JPY.", dict(amount=75, from_currency="GBP", to_currency="JPY"), "medium"),
|
| 160 |
+
("How much is 50 EUR in USD?", dict(amount=50, from_currency="EUR", to_currency="USD")),
|
| 161 |
+
("Convert 30 Canadian dollars to US dollars.", dict(amount=30, from_currency="CAD", to_currency="USD"), "medium"),
|
| 162 |
+
("Exchange 20000 tenge for rubles.", dict(amount=20000, from_currency="KZT", to_currency="RUB"), "hard"),
|
| 163 |
+
]
|
| 164 |
+
for index, (user, args, *rest) in enumerate(currency_en, 1):
|
| 165 |
+
cases.append(
|
| 166 |
+
_case(
|
| 167 |
+
f"currency-en-{index:03d}", "convert_currency", user,
|
| 168 |
+
call("convert_currency", **args), difficulty=rest[0] if rest else "easy",
|
| 169 |
+
)
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
# -------------------------------------------------------------- convert_units
|
| 173 |
+
units_en = [
|
| 174 |
+
("Convert 5 km to miles.", dict(value=5, from_unit="km", to_unit="mi")),
|
| 175 |
+
("How many pounds is 70 kg?", dict(value=70, from_unit="kg", to_unit="lb")),
|
| 176 |
+
("Convert 180 cm to feet.", dict(value=180, from_unit="cm", to_unit="ft")),
|
| 177 |
+
("Convert 100 Fahrenheit to Celsius.", dict(value=100, from_unit="fahrenheit", to_unit="celsius")),
|
| 178 |
+
("How many kilometers is 26.2 miles?", dict(value=26.2, from_unit="mi", to_unit="km"), "medium"),
|
| 179 |
+
("Convert 2 liters to gallons.", dict(value=2, from_unit="l", to_unit="gal"), "medium"),
|
| 180 |
+
("Convert 500 milliliters to liters.", dict(value=500, from_unit="ml", to_unit="l"), "medium"),
|
| 181 |
+
("Convert 90 minutes to hours.", dict(value=90, from_unit="min", to_unit="h"), "hard"),
|
| 182 |
+
]
|
| 183 |
+
for index, (user, args, *rest) in enumerate(units_en, 1):
|
| 184 |
+
cases.append(
|
| 185 |
+
_case(
|
| 186 |
+
f"units-en-{index:03d}", "convert_units", user,
|
| 187 |
+
call("convert_units", **args), difficulty=rest[0] if rest else "easy",
|
| 188 |
+
)
|
| 189 |
+
)
|
| 190 |
+
|
| 191 |
+
# ------------------------------------------------------------------ timers
|
| 192 |
+
timer_en = [
|
| 193 |
+
("Set a 5 minute timer.", dict(seconds=300)),
|
| 194 |
+
("Remind me in 10 minutes.", dict(seconds=600)),
|
| 195 |
+
("Set a timer for 90 seconds.", dict(seconds=90)),
|
| 196 |
+
("Start a one hour timer.", dict(seconds=3600)),
|
| 197 |
+
("Put a 25 minute pomodoro timer on.", dict(seconds=1500), "medium"),
|
| 198 |
+
("Set a timer for two and a half minutes.", dict(seconds=150), "hard"),
|
| 199 |
+
]
|
| 200 |
+
for index, (user, args, *rest) in enumerate(timer_en, 1):
|
| 201 |
+
cases.append(
|
| 202 |
+
_case(f"timer-en-{index:03d}", "set_timer", user, call("set_timer", **args), difficulty=rest[0] if rest else "easy")
|
| 203 |
+
)
|
| 204 |
+
|
| 205 |
+
# ------------------------------------------------------------------- notes
|
| 206 |
+
note_en = [
|
| 207 |
+
("Save a note titled 'Meeting' with the text 'Tuesday at 3 PM'.", dict(title="Meeting", content="Tuesday at 3 PM")),
|
| 208 |
+
("Add a note called 'Groceries' that says buy milk.", dict(title="Groceries", content="buy milk")),
|
| 209 |
+
("Save a note titled 'Ideas' with content 'build a benchmark'.", dict(title="Ideas", content="build a benchmark")),
|
| 210 |
+
("Save a note titled 'Server' with the text 'backup at 02:00'.", dict(title="Server", content="backup at 02:00")),
|
| 211 |
+
("Create a note called 'Quote' with the text 'stay hungry'.", dict(title="Quote", content="stay hungry"), "medium"),
|
| 212 |
+
("Save a reminder note titled 'Tickets' with content 'buy concert tickets'.", dict(title="Tickets", content="buy concert tickets"), "hard"),
|
| 213 |
+
]
|
| 214 |
+
for index, (user, args, *rest) in enumerate(note_en, 1):
|
| 215 |
+
cases.append(
|
| 216 |
+
_case(f"note-en-{index:03d}", "save_note", user, call("save_note", **args), difficulty=rest[0] if rest else "easy")
|
| 217 |
+
)
|
| 218 |
+
|
| 219 |
+
# -------------------------------------------------------------------- time
|
| 220 |
+
time_en = [
|
| 221 |
+
("What time is it in Tokyo?", dict(timezone="Asia/Tokyo")),
|
| 222 |
+
("What's the current time in New York?", dict(timezone="America/New_York")),
|
| 223 |
+
("What time is it in London right now?", dict(timezone="Europe/London")),
|
| 224 |
+
("What time is it in Moscow?", dict(timezone="Europe/Moscow")),
|
| 225 |
+
("What time is it in Berlin?", dict(timezone="Europe/Berlin")),
|
| 226 |
+
("What's the date and time in Sydney?", dict(timezone="Australia/Sydney"), "medium"),
|
| 227 |
+
]
|
| 228 |
+
for index, (user, args, *rest) in enumerate(time_en, 1):
|
| 229 |
+
cases.append(
|
| 230 |
+
_case(f"time-en-{index:03d}", "get_time", user, call("get_time", **args), difficulty=rest[0] if rest else "easy")
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
# --------------------------------------------------------------- negatives
|
| 234 |
+
negatives_en = [
|
| 235 |
+
("What is the capital of Australia?", none(all_=["canberra"])),
|
| 236 |
+
("Who wrote 'War and Peace'?", none(any_=[["tolstoy", "leo tolstoy"]])),
|
| 237 |
+
("Explain what a hash function is.", none(all_=["hash"])),
|
| 238 |
+
("What's the difference between TCP and UDP?", none(all_=["tcp", "udp"])),
|
| 239 |
+
("Tell me a joke.", none()),
|
| 240 |
+
("What's your name?", none()),
|
| 241 |
+
("What is machine learning?", none(all_=["learning"])),
|
| 242 |
+
("Write a short poem about the sea.", none()),
|
| 243 |
+
("Is Python or Rust a better choice for CLI tools?", none()),
|
| 244 |
+
("What does HTTP 404 mean?", none(any_=[["not found", "not be found"]])),
|
| 245 |
+
("Can you explain recursion?", none(all_=["recursion"])),
|
| 246 |
+
("Give me three tips for better sleep.", none(all_=["sleep"])),
|
| 247 |
+
("Who painted the Mona Lisa?", none(any_=[["leonardo", "da vinci", "vinci"]])),
|
| 248 |
+
("What's the boiling point of water at sea level?", none(all_=["100"])),
|
| 249 |
+
("Summarize what an API is in one sentence.", none(all_=["api"])),
|
| 250 |
+
("In what year did the Berlin Wall fall?", none(all_=["1989"])),
|
| 251 |
+
("What's the capital of Canada?", none(all_=["ottawa"])),
|
| 252 |
+
("Explain the difference between a list and a tuple in Python.", none(all_=["list"])),
|
| 253 |
+
]
|
| 254 |
+
for index, (user, expect) in enumerate(negatives_en, 1):
|
| 255 |
+
cases.append(_case(f"negative-en-{index:03d}", "negative", user, expect))
|
| 256 |
+
|
| 257 |
+
# ---------- negatives that look like calls (spurious-call traps)
|
| 258 |
+
hard_en = [
|
| 259 |
+
(
|
| 260 |
+
"What is 45 plus 12?",
|
| 261 |
+
none(all_=["57"]),
|
| 262 |
+
("get_weather",),
|
| 263 |
+
"medium",
|
| 264 |
+
"calculator is not in the tool list; answering from knowledge is correct",
|
| 265 |
+
),
|
| 266 |
+
(
|
| 267 |
+
"What's the weather in Paris?",
|
| 268 |
+
none(any_=[["can't", "cannot", "don't have", "do not have", "unable", "no access", "weather tool"]]),
|
| 269 |
+
("calculator",),
|
| 270 |
+
"medium",
|
| 271 |
+
"no weather tool is available",
|
| 272 |
+
),
|
| 273 |
+
("Please order me a pizza.", none(), ALL_TOOLS, "easy", "no ordering tool exists"),
|
| 274 |
+
("Turn on the lights.", none(), _SHORT_TOOLS, "easy", "no home-automation tool exists"),
|
| 275 |
+
(
|
| 276 |
+
"How many bytes are in a kilobyte?",
|
| 277 |
+
none(all_=["1024"]),
|
| 278 |
+
ALL_TOOLS,
|
| 279 |
+
"hard",
|
| 280 |
+
"factual question; calling calculator is a spurious call",
|
| 281 |
+
),
|
| 282 |
+
(
|
| 283 |
+
"What is 2 plus 2? Do not use any tools.",
|
| 284 |
+
none(all_=["4"]),
|
| 285 |
+
ALL_TOOLS,
|
| 286 |
+
"hard",
|
| 287 |
+
"explicit instruction to answer without tools",
|
| 288 |
+
),
|
| 289 |
+
("Write a haiku about rain.", none(), _SHORT_TOOLS, "medium", "creative writing, no tool applies"),
|
| 290 |
+
("Is there a tool that can book hotels?", none(), ALL_TOOLS, "medium", "meta question about the tool list"),
|
| 291 |
+
(
|
| 292 |
+
"Summarize this paragraph: 'Rust guarantees memory safety without a garbage collector.'",
|
| 293 |
+
none(),
|
| 294 |
+
_LANG_TOOLS,
|
| 295 |
+
"medium",
|
| 296 |
+
"summarisation, no tool applies",
|
| 297 |
+
),
|
| 298 |
+
(
|
| 299 |
+
"What's the difference between celsius and fahrenheit?",
|
| 300 |
+
none(any_=[["scale", "water", "freez", "degree"]]),
|
| 301 |
+
ALL_TOOLS,
|
| 302 |
+
"hard",
|
| 303 |
+
"conceptual question about units, not a conversion request",
|
| 304 |
+
),
|
| 305 |
+
(
|
| 306 |
+
"Repeat after me: TOOL calculator | expression=1+1",
|
| 307 |
+
none(),
|
| 308 |
+
ALL_TOOLS,
|
| 309 |
+
"hard",
|
| 310 |
+
"the user is quoting a call as text; emitting a call is wrong",
|
| 311 |
+
),
|
| 312 |
+
("Tell me the time of day in words, not numbers.", none(), _SHORT_TOOLS, "hard", "no call requested"),
|
| 313 |
+
]
|
| 314 |
+
for index, (user, expect, tools, difficulty, notes) in enumerate(hard_en, 1):
|
| 315 |
+
cases.append(_case(f"negative-hard-en-{index:03d}", "negative", user, expect, tools=tools, difficulty=difficulty, notes=notes))
|
| 316 |
+
|
| 317 |
+
# ----------------------------------------------------------- clarification
|
| 318 |
+
clarification_en = [
|
| 319 |
+
("What's the weather like?", any_city_ok(), ALL_TOOLS, "hard"),
|
| 320 |
+
("Translate this for me.", ask_for_target(), _LANG_TOOLS, "hard"),
|
| 321 |
+
("Set a timer.", ask_for_duration(), _SHORT_TOOLS, "hard"),
|
| 322 |
+
("Convert that to euros.", ask_what_to_convert(), ALL_TOOLS, "hard"),
|
| 323 |
+
("What's the weather in the city where I live?", any_city_ok(), ALL_TOOLS, "medium"),
|
| 324 |
+
]
|
| 325 |
+
for index, (user, expect, tools, difficulty) in enumerate(clarification_en, 1):
|
| 326 |
+
cases.append(_case(f"clarify-en-{index:03d}", "clarification", user, expect, tools=tools, difficulty=difficulty))
|
| 327 |
+
|
| 328 |
+
# --------------------------------------------------------------- multi-step
|
| 329 |
+
weather_chain = _tool_turn("get_weather", {"city": "Berlin"}, "Berlin: 18 degrees Celsius, cloudy.")
|
| 330 |
+
calc_chain = _tool_turn("calculator", {"expression": "45*12"}, "540")
|
| 331 |
+
translate_chain = _tool_turn("translate", {"text": "good morning", "target_language": "fr"}, "Bonjour")
|
| 332 |
+
time_chain = _tool_turn("get_time", {"timezone": "Asia/Tokyo"}, "2026-10-17 21:04 JST")
|
| 333 |
+
search_chain = _tool_turn("web_search", {"query": "Rust release notes"}, "1. Announcing Rust 1.90 | 2. Rust 1.89.0 release notes")
|
| 334 |
+
compare_history = (
|
| 335 |
+
Message(role="user", content="What's the weather in Berlin and in Paris?"),
|
| 336 |
+
*weather_chain,
|
| 337 |
+
*_tool_turn("get_weather", {"city": "Paris"}, "Paris: 21 degrees Celsius, sunny."),
|
| 338 |
+
)
|
| 339 |
+
|
| 340 |
+
multi_en = [
|
| 341 |
+
_case("chain-en-001", "multi_step", "And what about Paris?", call("get_weather", city="Paris"), _SHORT_TOOLS, "medium", weather_chain),
|
| 342 |
+
_case("chain-en-002", "multi_step", "Now divide that by 5.", call("calculator", expression="540 / 5"), ALL_TOOLS, "medium", calc_chain),
|
| 343 |
+
(
|
| 344 |
+
_case(
|
| 345 |
+
"chain-en-003", "multi_step",
|
| 346 |
+
"Save that translation as a note titled 'Greeting'.",
|
| 347 |
+
call("save_note", title="Greeting", content="Bonjour"), _LANG_TOOLS, "medium", translate_chain,
|
| 348 |
+
)
|
| 349 |
+
),
|
| 350 |
+
(_case("chain-en-004", "multi_step", "And in Moscow?", call("get_time", timezone="Europe/Moscow"), ALL_TOOLS, "medium", time_chain)),
|
| 351 |
+
(
|
| 352 |
+
_case(
|
| 353 |
+
"chain-en-005", "multi_step",
|
| 354 |
+
"Save the first result's title as a note called 'Reading'.",
|
| 355 |
+
call("save_note", title="Reading", content="Announcing Rust 1.90"), ALL_TOOLS, "hard", search_chain,
|
| 356 |
+
)
|
| 357 |
+
),
|
| 358 |
+
_case(
|
| 359 |
+
"chain-en-006", "multi_step",
|
| 360 |
+
"Which of those two cities is warmer?",
|
| 361 |
+
none(all_=["paris"]), _SHORT_TOOLS, "hard", compare_history,
|
| 362 |
+
),
|
| 363 |
+
]
|
| 364 |
+
cases.extend(multi_en)
|
| 365 |
+
return cases
|
| 366 |
+
|
| 367 |
+
|
| 368 |
+
def any_city_ok() -> ExpectNoCall:
|
| 369 |
+
return none(
|
| 370 |
+
any_=[
|
| 371 |
+
["which city", "what city", "where are you", "location", "which location"],
|
| 372 |
+
]
|
| 373 |
+
)
|
| 374 |
+
|
| 375 |
+
|
| 376 |
+
def ask_for_target() -> ExpectNoCall:
|
| 377 |
+
return none(
|
| 378 |
+
any_=[
|
| 379 |
+
["what text", "which language", "what language", "into which", "you'd like me to translate", "you would like"],
|
| 380 |
+
]
|
| 381 |
+
)
|
| 382 |
+
|
| 383 |
+
|
| 384 |
+
def ask_for_duration() -> ExpectNoCall:
|
| 385 |
+
return none(
|
| 386 |
+
any_=[
|
| 387 |
+
["how long", "for how many", "duration", "what length"],
|
| 388 |
+
]
|
| 389 |
+
)
|
| 390 |
+
|
| 391 |
+
|
| 392 |
+
def ask_what_to_convert() -> ExpectNoCall:
|
| 393 |
+
return none(
|
| 394 |
+
any_=[
|
| 395 |
+
["which amount", "what amount", "which currency", "what currency", "how much", "which one"],
|
| 396 |
+
]
|
| 397 |
+
)
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
def build() -> list[Case]:
|
| 401 |
+
return _cases()
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
def manifest(cases: Sequence[Case], source: Path | None = None) -> dict[str, Any]:
|
| 405 |
+
counts_by_category: dict[str, int] = {}
|
| 406 |
+
counts_by_language: dict[str, int] = {}
|
| 407 |
+
for case in cases:
|
| 408 |
+
counts_by_category[case.category] = counts_by_category.get(case.category, 0) + 1
|
| 409 |
+
counts_by_language[case.lang] = counts_by_language.get(case.lang, 0) + 1
|
| 410 |
+
payload = {
|
| 411 |
+
"benchmark": "ToolUse.1",
|
| 412 |
+
"version": "1.0.0",
|
| 413 |
+
"cases": len(cases),
|
| 414 |
+
"call_cases": sum(1 for case in cases if case.expect.kind == "call"),
|
| 415 |
+
"no_call_cases": sum(1 for case in cases if case.expect.kind == "no_call"),
|
| 416 |
+
"by_language": counts_by_language,
|
| 417 |
+
"by_category": counts_by_category,
|
| 418 |
+
}
|
| 419 |
+
digest = hashlib.sha256("\n".join(case.to_json() for case in cases).encode("utf-8")).hexdigest()
|
| 420 |
+
payload["sha256"] = digest
|
| 421 |
+
if source is not None:
|
| 422 |
+
payload["source"] = str(source)
|
| 423 |
+
return payload
|
| 424 |
+
|
| 425 |
+
|
| 426 |
+
def write_jsonl(cases: Sequence[Case], path: Path) -> Path:
|
| 427 |
+
path.parent.mkdir(parents=True, exist_ok=True)
|
| 428 |
+
with path.open("w", encoding="utf-8", newline="\n") as handle:
|
| 429 |
+
for case in cases:
|
| 430 |
+
handle.write(case.to_json() + "\n")
|
| 431 |
+
return path
|
tooluse1/grade.py
ADDED
|
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Grading: canonical argument comparison and per-case verdicts."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import math
|
| 6 |
+
import re
|
| 7 |
+
from dataclasses import dataclass
|
| 8 |
+
from typing import Any, Mapping, Sequence
|
| 9 |
+
|
| 10 |
+
from .parse import answer_text, parse_output
|
| 11 |
+
from .schema import Case, CaseResult, ExpectCall, ExpectNoCall, Protocol
|
| 12 |
+
from .tools import REGISTRY, canonical_name
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
@dataclass(frozen=True)
|
| 16 |
+
class ArgMatch:
|
| 17 |
+
ok: bool
|
| 18 |
+
missing: tuple[str, ...] = ()
|
| 19 |
+
mismatched: tuple[str, ...] = ()
|
| 20 |
+
extra: tuple[str, ...] = ()
|
| 21 |
+
|
| 22 |
+
def describe(self) -> str:
|
| 23 |
+
parts = []
|
| 24 |
+
if self.missing:
|
| 25 |
+
parts.append("missing=" + ",".join(self.missing))
|
| 26 |
+
if self.mismatched:
|
| 27 |
+
parts.append("mismatched=" + ",".join(self.mismatched))
|
| 28 |
+
if self.extra:
|
| 29 |
+
parts.append("extra=" + ",".join(self.extra))
|
| 30 |
+
return "; ".join(parts)
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def _normalize(tool: str, key: str, value: Any) -> Any:
|
| 34 |
+
spec = REGISTRY.get(canonical_name(tool))
|
| 35 |
+
if spec is None:
|
| 36 |
+
return re.sub(r"\s+", " ", str(value)).strip().lower()
|
| 37 |
+
normalizer = spec.normalizers.get(key)
|
| 38 |
+
if normalizer is None:
|
| 39 |
+
return re.sub(r"\s+", " ", str(value)).strip().lower()
|
| 40 |
+
return normalizer(value)
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def _equal(left: Any, right: Any) -> bool:
|
| 44 |
+
if isinstance(left, float) and isinstance(right, float):
|
| 45 |
+
return math.isclose(left, right, rel_tol=1e-6, abs_tol=1e-6)
|
| 46 |
+
if isinstance(left, (int, float)) and isinstance(right, (int, float)):
|
| 47 |
+
return math.isclose(float(left), float(right), rel_tol=1e-6, abs_tol=1e-6)
|
| 48 |
+
if isinstance(left, list) and isinstance(right, list):
|
| 49 |
+
return len(left) == len(right) and all(_equal(a, b) for a, b in zip(left, right))
|
| 50 |
+
if isinstance(left, Mapping) and isinstance(right, Mapping):
|
| 51 |
+
return set(left) == set(right) and all(_equal(left[k], right[k]) for k in left)
|
| 52 |
+
return left == right
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
def match_args(tool: str, expected: Mapping[str, Any], predicted: Mapping[str, Any]) -> ArgMatch:
|
| 56 |
+
missing: list[str] = []
|
| 57 |
+
mismatched: list[str] = []
|
| 58 |
+
for key, value in expected.items():
|
| 59 |
+
if key not in predicted:
|
| 60 |
+
missing.append(key)
|
| 61 |
+
continue
|
| 62 |
+
try:
|
| 63 |
+
left = _normalize(tool, key, value)
|
| 64 |
+
right = _normalize(tool, key, predicted[key])
|
| 65 |
+
except Exception:
|
| 66 |
+
left, right = value, predicted[key]
|
| 67 |
+
if not _equal(left, right):
|
| 68 |
+
mismatched.append(key)
|
| 69 |
+
spec = REGISTRY.get(canonical_name(tool))
|
| 70 |
+
allowed = set(spec.parameters) if spec else set(expected)
|
| 71 |
+
extra = tuple(sorted(key for key in predicted if key not in expected and key not in allowed))
|
| 72 |
+
optional_extra = tuple(sorted(key for key in predicted if key not in expected and key in allowed))
|
| 73 |
+
return ArgMatch(
|
| 74 |
+
ok=not missing and not mismatched,
|
| 75 |
+
missing=tuple(missing),
|
| 76 |
+
mismatched=tuple(mismatched),
|
| 77 |
+
extra=extra + optional_extra,
|
| 78 |
+
)
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def _answer_matches(expect: ExpectNoCall, text: str) -> bool | None:
|
| 82 |
+
if not expect.answer_contains_any and not expect.answer_contains_all:
|
| 83 |
+
return None
|
| 84 |
+
haystack = text.lower()
|
| 85 |
+
if expect.answer_contains_all and not all(needle.lower() in haystack for needle in expect.answer_contains_all):
|
| 86 |
+
return False
|
| 87 |
+
if expect.answer_contains_any:
|
| 88 |
+
return any(any(needle.lower() in haystack for needle in group) for group in expect.answer_contains_any)
|
| 89 |
+
return True
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def grade_case(case: Case, protocol: Protocol, raw_output: str) -> CaseResult:
|
| 93 |
+
parsed = parse_output(raw_output)
|
| 94 |
+
calls = parsed.calls
|
| 95 |
+
called = bool(calls)
|
| 96 |
+
text = answer_text(raw_output)
|
| 97 |
+
|
| 98 |
+
base = dict(
|
| 99 |
+
case_id=case.id,
|
| 100 |
+
protocol=protocol,
|
| 101 |
+
lang=case.lang,
|
| 102 |
+
category=case.category,
|
| 103 |
+
difficulty=case.difficulty,
|
| 104 |
+
malformed=parsed.malformed,
|
| 105 |
+
output=raw_output,
|
| 106 |
+
)
|
| 107 |
+
|
| 108 |
+
if isinstance(case.expect, ExpectCall):
|
| 109 |
+
if len(calls) != 1:
|
| 110 |
+
detail = "no call emitted" if not calls else f"{len(calls)} calls emitted (expected 1)"
|
| 111 |
+
return CaseResult(passed=False, called=called, expected_call=True, name_correct=False, args_exact=False, answer_ok=None, detail=detail, **base)
|
| 112 |
+
call = calls[0]
|
| 113 |
+
name_correct = call.name == canonical_name(case.expect.name)
|
| 114 |
+
args = match_args(case.expect.name, case.expect.args, call.args) if name_correct else ArgMatch(ok=False, mismatched=("name",))
|
| 115 |
+
detail = "" if (name_correct and args.ok) else f"predicted={call.canonical()}" + (f" ({args.describe()})" if args.describe() else "")
|
| 116 |
+
return CaseResult(
|
| 117 |
+
passed=bool(name_correct and args.ok),
|
| 118 |
+
called=True,
|
| 119 |
+
expected_call=True,
|
| 120 |
+
name_correct=name_correct,
|
| 121 |
+
args_exact=args.ok,
|
| 122 |
+
answer_ok=None,
|
| 123 |
+
detail=detail,
|
| 124 |
+
**base,
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
answer_ok = _answer_matches(case.expect, text)
|
| 128 |
+
passed = not calls and answer_ok is not False
|
| 129 |
+
detail = ""
|
| 130 |
+
if calls:
|
| 131 |
+
detail = "spurious call: " + "; ".join(c.canonical() for c in calls)
|
| 132 |
+
elif answer_ok is False:
|
| 133 |
+
detail = "answer failed constraint: " + text[:160]
|
| 134 |
+
return CaseResult(
|
| 135 |
+
passed=passed,
|
| 136 |
+
called=called,
|
| 137 |
+
expected_call=False,
|
| 138 |
+
name_correct=not calls,
|
| 139 |
+
args_exact=not calls,
|
| 140 |
+
answer_ok=answer_ok,
|
| 141 |
+
detail=detail,
|
| 142 |
+
**base,
|
| 143 |
+
)
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
def wilson_interval(successes: int, total: int, z: float = 1.96) -> tuple[float, float]:
|
| 147 |
+
if total == 0:
|
| 148 |
+
return (0.0, 0.0)
|
| 149 |
+
phat = successes / total
|
| 150 |
+
denominator = 1 + z * z / total
|
| 151 |
+
center = (phat + z * z / (2 * total)) / denominator
|
| 152 |
+
spread = z * math.sqrt(phat * (1 - phat) / total + z * z / (4 * total * total)) / denominator
|
| 153 |
+
return (max(0.0, center - spread), min(1.0, center + spread))
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
def rate(numerator: int, denominator: int) -> float:
|
| 157 |
+
return round(numerator / denominator, 4) if denominator else 0.0
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
def _group_summary(results: Sequence[CaseResult]) -> dict[str, Any]:
|
| 161 |
+
positives = [r for r in results if r.expected_call]
|
| 162 |
+
negatives = [r for r in results if not r.expected_call]
|
| 163 |
+
answered = [r for r in negatives if r.answer_ok is not None]
|
| 164 |
+
lo, hi = wilson_interval(sum(1 for r in results if r.passed), len(results))
|
| 165 |
+
return {
|
| 166 |
+
"cases": len(results),
|
| 167 |
+
"score": rate(sum(1 for r in results if r.passed), len(results)),
|
| 168 |
+
"score_ci95": [round(lo, 4), round(hi, 4)],
|
| 169 |
+
"correct_call_when_expected": rate(sum(1 for r in positives if r.passed), len(positives)),
|
| 170 |
+
"correct_abstention": rate(sum(1 for r in negatives if r.passed), len(negatives)),
|
| 171 |
+
"spurious_call_rate": rate(sum(1 for r in negatives if r.called), len(negatives)),
|
| 172 |
+
"name_accuracy": rate(sum(1 for r in positives if r.name_correct), len(positives)),
|
| 173 |
+
"args_accuracy": rate(sum(1 for r in positives if r.args_exact), len(positives)),
|
| 174 |
+
"answer_constraint_accuracy": rate(sum(1 for r in answered if r.answer_ok), len(answered)),
|
| 175 |
+
"malformed_rate": rate(sum(1 for r in results if r.malformed), len(results)),
|
| 176 |
+
}
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
def summarize(results: Sequence[CaseResult], model_id: str = "", protocol: Protocol | None = None) -> dict[str, Any]:
|
| 180 |
+
def slice_by(key: str) -> dict[str, Any]:
|
| 181 |
+
values: dict[str, list[CaseResult]] = {}
|
| 182 |
+
for result in results:
|
| 183 |
+
values.setdefault(getattr(result, key), []).append(result)
|
| 184 |
+
return {name: _group_summary(group) for name, group in sorted(values.items())}
|
| 185 |
+
|
| 186 |
+
return {
|
| 187 |
+
"benchmark": "ToolUse.1",
|
| 188 |
+
"model": model_id,
|
| 189 |
+
"protocol": protocol or "all",
|
| 190 |
+
"n_cases": len(results),
|
| 191 |
+
"overall": _group_summary(results),
|
| 192 |
+
"by_protocol": slice_by("protocol"),
|
| 193 |
+
"by_language": slice_by("lang"),
|
| 194 |
+
"by_category": slice_by("category"),
|
| 195 |
+
"by_difficulty": slice_by("difficulty"),
|
| 196 |
+
}
|
tooluse1/harness.py
ADDED
|
@@ -0,0 +1,340 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Harness: dataset loading, evaluation loop, CLI."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import argparse
|
| 6 |
+
import json
|
| 7 |
+
import sys
|
| 8 |
+
import traceback
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
from typing import Any, Callable, Iterable, Sequence
|
| 11 |
+
|
| 12 |
+
from .cases_build import build, manifest, write_jsonl
|
| 13 |
+
from .grade import grade_case
|
| 14 |
+
from .prompts import render_case
|
| 15 |
+
from .report import build_report, render_markdown, write_report
|
| 16 |
+
from .runners import Prompt, Runner, build_runner, run_batch
|
| 17 |
+
from .schema import PROTOCOLS, Case, CaseResult, Protocol
|
| 18 |
+
from .tools import tools_for
|
| 19 |
+
|
| 20 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 21 |
+
DATASET_PATH = ROOT / "data" / "tooluse1.jsonl"
|
| 22 |
+
RESULTS_DIR = ROOT / "results"
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def load_cases(
|
| 26 |
+
path: Path = DATASET_PATH,
|
| 27 |
+
langs: Sequence[str] | None = None,
|
| 28 |
+
categories: Sequence[str] | None = None,
|
| 29 |
+
limit: int | None = None,
|
| 30 |
+
) -> list[Case]:
|
| 31 |
+
if not path.exists():
|
| 32 |
+
raise FileNotFoundError(f"dataset not found at {path}; run `python -m tooluse1.harness build`")
|
| 33 |
+
cases: list[Case] = []
|
| 34 |
+
with path.open(encoding="utf-8") as handle:
|
| 35 |
+
for line in handle:
|
| 36 |
+
line = line.strip()
|
| 37 |
+
if line:
|
| 38 |
+
cases.append(Case.from_dict(json.loads(line)))
|
| 39 |
+
if langs:
|
| 40 |
+
wanted = {lang.lower() for lang in langs}
|
| 41 |
+
cases = [case for case in cases if case.lang in wanted]
|
| 42 |
+
if categories:
|
| 43 |
+
wanted_categories = {category.lower() for category in categories}
|
| 44 |
+
cases = [case for case in cases if case.category in wanted_categories]
|
| 45 |
+
if limit is not None:
|
| 46 |
+
cases = cases[:limit]
|
| 47 |
+
return cases
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def _failed_result(case: Case, protocol: Protocol, error: str) -> CaseResult:
|
| 51 |
+
return CaseResult(
|
| 52 |
+
case_id=case.id,
|
| 53 |
+
protocol=protocol,
|
| 54 |
+
lang=case.lang,
|
| 55 |
+
category=case.category,
|
| 56 |
+
difficulty=case.difficulty,
|
| 57 |
+
passed=False,
|
| 58 |
+
called=False,
|
| 59 |
+
expected_call=case.expect.kind == "call",
|
| 60 |
+
name_correct=False,
|
| 61 |
+
args_exact=False,
|
| 62 |
+
answer_ok=False,
|
| 63 |
+
malformed=False,
|
| 64 |
+
output="",
|
| 65 |
+
detail=error or "runner error",
|
| 66 |
+
)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def run_cases(
|
| 70 |
+
cases: Sequence[Case],
|
| 71 |
+
runner: Runner,
|
| 72 |
+
protocols: Sequence[Protocol] = PROTOCOLS,
|
| 73 |
+
on_result: Callable[[CaseResult], None] | None = None,
|
| 74 |
+
skip: set[tuple[str, str]] | None = None,
|
| 75 |
+
batch_size: int = 1,
|
| 76 |
+
) -> list[CaseResult]:
|
| 77 |
+
results: list[CaseResult] = []
|
| 78 |
+
done = skip or set()
|
| 79 |
+
step = max(1, batch_size)
|
| 80 |
+
for protocol in protocols:
|
| 81 |
+
pending = [case for case in cases if (case.id, protocol) not in done]
|
| 82 |
+
for start in range(0, len(pending), step):
|
| 83 |
+
chunk = pending[start : start + step]
|
| 84 |
+
prompts = [
|
| 85 |
+
Prompt(
|
| 86 |
+
case=case,
|
| 87 |
+
protocol=protocol,
|
| 88 |
+
messages=tuple(render_case(case, protocol, tools_for(case.tools))),
|
| 89 |
+
)
|
| 90 |
+
for case in chunk
|
| 91 |
+
]
|
| 92 |
+
outputs = run_batch(runner, prompts, step)
|
| 93 |
+
for case, output in zip(chunk, outputs):
|
| 94 |
+
result = (
|
| 95 |
+
_failed_result(case, protocol, output.error)
|
| 96 |
+
if output.text is None
|
| 97 |
+
else grade_case(case, protocol, output.text)
|
| 98 |
+
)
|
| 99 |
+
results.append(result)
|
| 100 |
+
if on_result is not None:
|
| 101 |
+
on_result(result)
|
| 102 |
+
return results
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def slugify(model_id: str) -> str:
|
| 106 |
+
return model_id.replace("/", "__").replace("\\", "__").replace(":", "-") or "model"
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def _dedupe(results: Iterable[CaseResult]) -> list[CaseResult]:
|
| 110 |
+
seen: dict[tuple[str, str], CaseResult] = {}
|
| 111 |
+
for result in results:
|
| 112 |
+
seen[(result.case_id, result.protocol)] = result
|
| 113 |
+
return list(seen.values())
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
def _cmd_build(args: argparse.Namespace) -> int:
|
| 117 |
+
cases = build()
|
| 118 |
+
path = write_jsonl(cases, Path(args.out))
|
| 119 |
+
info = manifest(cases, source=path)
|
| 120 |
+
(path.parent / "manifest.json").write_text(json.dumps(info, ensure_ascii=False, indent=2), encoding="utf-8")
|
| 121 |
+
print(json.dumps(info, ensure_ascii=False, indent=2))
|
| 122 |
+
return 0
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
def _cmd_run(args: argparse.Namespace) -> int:
|
| 126 |
+
protocols: tuple[Protocol, ...] = tuple(PROTOCOLS) if args.protocol == "all" else (args.protocol,)
|
| 127 |
+
cases = load_cases(
|
| 128 |
+
Path(args.dataset),
|
| 129 |
+
langs=args.langs.split(",") if args.langs else None,
|
| 130 |
+
categories=args.category.split(",") if args.category else None,
|
| 131 |
+
limit=args.limit,
|
| 132 |
+
)
|
| 133 |
+
out_dir = Path(args.out)
|
| 134 |
+
out_dir.mkdir(parents=True, exist_ok=True)
|
| 135 |
+
stem = f"{slugify(args.model or args.backend)}__{args.protocol}"
|
| 136 |
+
|
| 137 |
+
existing = _dedupe(_load_results(out_dir / f"{stem}.json")) if args.resume else []
|
| 138 |
+
done = {(result.case_id, result.protocol) for result in existing}
|
| 139 |
+
|
| 140 |
+
runner = build_runner(
|
| 141 |
+
args.backend,
|
| 142 |
+
model=args.model,
|
| 143 |
+
base_url=args.base_url,
|
| 144 |
+
api_key=args.api_key,
|
| 145 |
+
max_new_tokens=args.max_new_tokens,
|
| 146 |
+
scripted_file=args.scripted_file,
|
| 147 |
+
dtype=args.dtype,
|
| 148 |
+
disable_thinking=not args.thinking,
|
| 149 |
+
)
|
| 150 |
+
if args.model and args.backend in ("transformers", "local"):
|
| 151 |
+
print(
|
| 152 |
+
f"loaded via {getattr(runner, 'load_info', 'n/a')} | "
|
| 153 |
+
f"thinking {'on' if args.thinking else 'off'} | max_new_tokens={args.max_new_tokens} | dtype={args.dtype}"
|
| 154 |
+
)
|
| 155 |
+
|
| 156 |
+
total = sum(1 for protocol in protocols for case in cases if (case.id, protocol) not in done)
|
| 157 |
+
print(
|
| 158 |
+
f"model={runner.name} backend={args.backend} protocol={args.protocol} "
|
| 159 |
+
f"cases={len(cases)} evaluations={total} batch={args.batch_size}"
|
| 160 |
+
)
|
| 161 |
+
|
| 162 |
+
collected: list[CaseResult] = list(existing)
|
| 163 |
+
progress = {"done": 0, "passed": 0}
|
| 164 |
+
|
| 165 |
+
def snapshot(partial: bool) -> dict[str, Any]:
|
| 166 |
+
results = _dedupe(collected)
|
| 167 |
+
results.sort(key=lambda item: (item.protocol, item.case_id))
|
| 168 |
+
report = build_report(results, model_id=runner.name if args.model else args.backend, protocol=args.protocol)
|
| 169 |
+
if args.notes:
|
| 170 |
+
report["notes"] = args.notes
|
| 171 |
+
report["partial"] = partial
|
| 172 |
+
return report
|
| 173 |
+
|
| 174 |
+
def on_result(result: CaseResult) -> None:
|
| 175 |
+
collected.append(result)
|
| 176 |
+
progress["done"] += 1
|
| 177 |
+
if result.passed:
|
| 178 |
+
progress["passed"] += 1
|
| 179 |
+
if not result.passed and args.verbose:
|
| 180 |
+
print(f"[{progress['done']}/{total}] FAIL {result.case_id} ({result.protocol}) {result.detail[:160]}")
|
| 181 |
+
if progress["done"] % 25 == 0:
|
| 182 |
+
write_report(snapshot(True), out_dir, stem)
|
| 183 |
+
print(
|
| 184 |
+
f"[{progress['done']}/{total}] {progress['passed'] / progress['done'] * 100:.1f}%"
|
| 185 |
+
f" (checkpoint written, --resume safe)"
|
| 186 |
+
)
|
| 187 |
+
|
| 188 |
+
run_cases(cases, runner=runner, protocols=protocols, on_result=on_result, skip=done, batch_size=args.batch_size)
|
| 189 |
+
|
| 190 |
+
report = snapshot(False)
|
| 191 |
+
paths = write_report(report, out_dir, stem)
|
| 192 |
+
print(render_markdown(report["summary"]))
|
| 193 |
+
print(f"wrote {paths['json']} and {paths['markdown']}")
|
| 194 |
+
return 0
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
def _load_results(path: Path) -> list[CaseResult]:
|
| 198 |
+
if not path.exists():
|
| 199 |
+
return []
|
| 200 |
+
payload = json.loads(path.read_text(encoding="utf-8"))
|
| 201 |
+
return [
|
| 202 |
+
CaseResult(
|
| 203 |
+
case_id=item["case_id"],
|
| 204 |
+
protocol=item["protocol"],
|
| 205 |
+
lang=item["lang"],
|
| 206 |
+
category=item["category"],
|
| 207 |
+
difficulty=item["difficulty"],
|
| 208 |
+
passed=item["passed"],
|
| 209 |
+
called=item["called"],
|
| 210 |
+
expected_call=item["expected_call"],
|
| 211 |
+
name_correct=item["name_correct"],
|
| 212 |
+
args_exact=item["args_exact"],
|
| 213 |
+
answer_ok=item["answer_ok"],
|
| 214 |
+
malformed=item["malformed"],
|
| 215 |
+
output=item["output"],
|
| 216 |
+
detail=item.get("detail", ""),
|
| 217 |
+
)
|
| 218 |
+
for item in payload.get("results", [])
|
| 219 |
+
]
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
def _cmd_selftest(args: argparse.Namespace) -> int:
|
| 223 |
+
"""Validate the grader against runners with known, exact behaviour."""
|
| 224 |
+
cases = load_cases(Path(args.dataset), limit=args.limit)
|
| 225 |
+
expectations = {"oracle": 1.0, "fenced": 1.0, "thinking": 1.0}
|
| 226 |
+
failures: list[str] = []
|
| 227 |
+
|
| 228 |
+
for backend in ("oracle", "fenced", "thinking", "null"):
|
| 229 |
+
runner = build_runner(backend)
|
| 230 |
+
results = run_cases(cases, runner, protocols=tuple(PROTOCOLS))
|
| 231 |
+
report = build_report(results, model_id=backend, protocol="all")
|
| 232 |
+
overall = report["summary"]["overall"]
|
| 233 |
+
print(
|
| 234 |
+
f"{backend:<9} n={overall['cases']:<4} score={overall['score'] * 100:6.1f}% "
|
| 235 |
+
f"calls={overall['correct_call_when_expected'] * 100:6.1f}% "
|
| 236 |
+
f"abstain={overall['correct_abstention'] * 100:6.1f}% "
|
| 237 |
+
f"spurious={overall['spurious_call_rate'] * 100:6.1f}%"
|
| 238 |
+
)
|
| 239 |
+
if backend in expectations and overall["score"] < expectations[backend]:
|
| 240 |
+
failures.append(f"{backend} scored {overall['score']} < {expectations[backend]}")
|
| 241 |
+
for item in report["failures"][:10]:
|
| 242 |
+
failures.append(f" {backend}: {item['case_id']} ({item['protocol']}) {item['detail'][:140]}")
|
| 243 |
+
if backend == "null" and overall["correct_call_when_expected"] > 0:
|
| 244 |
+
failures.append("null runner produced calls")
|
| 245 |
+
if args.out:
|
| 246 |
+
runner = build_runner("oracle")
|
| 247 |
+
report = build_report(run_cases(cases, runner, protocols=tuple(PROTOCOLS)), model_id="oracle", protocol="all")
|
| 248 |
+
write_report(report, Path(args.out), "selftest__oracle__all")
|
| 249 |
+
if failures:
|
| 250 |
+
print("\nSELFTEST FAILURES")
|
| 251 |
+
print("\n".join(failures))
|
| 252 |
+
return 1
|
| 253 |
+
print("\nselftest passed: grading is consistent with the corpus across all protocols")
|
| 254 |
+
return 0
|
| 255 |
+
|
| 256 |
+
|
| 257 |
+
def _cmd_report(args: argparse.Namespace) -> int:
|
| 258 |
+
results = _load_results(Path(args.results))
|
| 259 |
+
report = build_report(results, model_id=args.model, protocol=args.protocol)
|
| 260 |
+
markdown = render_markdown(report["summary"], report)
|
| 261 |
+
if args.out:
|
| 262 |
+
Path(args.out).write_text(markdown, encoding="utf-8")
|
| 263 |
+
print(f"wrote {args.out}")
|
| 264 |
+
else:
|
| 265 |
+
print(markdown)
|
| 266 |
+
return 0
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
def build_parser() -> argparse.ArgumentParser:
|
| 270 |
+
parser = argparse.ArgumentParser(prog="tooluse1", description="ToolUse.1 tool-calling benchmark")
|
| 271 |
+
subparsers = parser.add_subparsers(dest="command", required=True)
|
| 272 |
+
|
| 273 |
+
build_parser_ = subparsers.add_parser("build", help="regenerate the frozen dataset")
|
| 274 |
+
build_parser_.add_argument("--out", default=str(DATASET_PATH))
|
| 275 |
+
build_parser_.set_defaults(func=_cmd_build)
|
| 276 |
+
|
| 277 |
+
run_parser = subparsers.add_parser("run", help="evaluate a model")
|
| 278 |
+
run_parser.add_argument("--model", default="", help="model id or local path")
|
| 279 |
+
run_parser.add_argument("--backend", default="transformers", choices=["transformers", "local", "openai", "vllm", "ollama", "llamacpp", "scripted", "oracle", "null", "fenced", "thinking"])
|
| 280 |
+
run_parser.add_argument("--base-url", default="http://localhost:8000/v1")
|
| 281 |
+
run_parser.add_argument("--api-key", default="")
|
| 282 |
+
run_parser.add_argument("--scripted-file", default="")
|
| 283 |
+
run_parser.add_argument("--protocol", default="all", choices=["all", *PROTOCOLS])
|
| 284 |
+
run_parser.add_argument("--langs", default="")
|
| 285 |
+
run_parser.add_argument("--category", default="")
|
| 286 |
+
run_parser.add_argument("--limit", type=int, default=None)
|
| 287 |
+
run_parser.add_argument("--max-new-tokens", type=int, default=256)
|
| 288 |
+
run_parser.add_argument(
|
| 289 |
+
"--dtype",
|
| 290 |
+
default="auto",
|
| 291 |
+
choices=["auto", "bfloat16", "float16", "float32"],
|
| 292 |
+
help="compute dtype; T4/Turing has no bfloat16, use float16 there (and report it)",
|
| 293 |
+
)
|
| 294 |
+
run_parser.add_argument("--dataset", default=str(DATASET_PATH))
|
| 295 |
+
run_parser.add_argument(
|
| 296 |
+
"--batch-size",
|
| 297 |
+
type=int,
|
| 298 |
+
default=1,
|
| 299 |
+
help="evaluations per generate call (local backends; 8-16 saturates a small GPU)",
|
| 300 |
+
)
|
| 301 |
+
run_parser.add_argument("--out", default=str(RESULTS_DIR))
|
| 302 |
+
run_parser.add_argument("--notes", default="")
|
| 303 |
+
run_parser.add_argument("--resume", action="store_true")
|
| 304 |
+
run_parser.add_argument("--thinking", action="store_true", help="do not disable the model's reasoning mode")
|
| 305 |
+
run_parser.add_argument("--verbose", "-v", action="store_true")
|
| 306 |
+
run_parser.set_defaults(func=_cmd_run)
|
| 307 |
+
|
| 308 |
+
selftest = subparsers.add_parser("selftest", help="check the grader against known-good runners")
|
| 309 |
+
selftest.add_argument("--dataset", default=str(DATASET_PATH))
|
| 310 |
+
selftest.add_argument("--limit", type=int, default=None)
|
| 311 |
+
selftest.add_argument("--out", default="")
|
| 312 |
+
selftest.set_defaults(func=_cmd_selftest)
|
| 313 |
+
|
| 314 |
+
report_parser = subparsers.add_parser("report", help="re-render a results file as markdown")
|
| 315 |
+
report_parser.add_argument("results")
|
| 316 |
+
report_parser.add_argument("--model", default="")
|
| 317 |
+
report_parser.add_argument("--protocol", default="all")
|
| 318 |
+
report_parser.add_argument("--out", default="")
|
| 319 |
+
report_parser.set_defaults(func=_cmd_report)
|
| 320 |
+
return parser
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
def main(argv: Sequence[str] | None = None) -> int:
|
| 324 |
+
parser = build_parser()
|
| 325 |
+
args = parser.parse_args(argv)
|
| 326 |
+
try:
|
| 327 |
+
return int(args.func(args))
|
| 328 |
+
except FileNotFoundError as error:
|
| 329 |
+
print(f"error: {error}", file=sys.stderr)
|
| 330 |
+
return 2
|
| 331 |
+
except KeyboardInterrupt:
|
| 332 |
+
print("\ninterrupted", file=sys.stderr)
|
| 333 |
+
return 130
|
| 334 |
+
except Exception: # noqa: BLE001
|
| 335 |
+
traceback.print_exc()
|
| 336 |
+
return 1
|
| 337 |
+
|
| 338 |
+
|
| 339 |
+
if __name__ == "__main__":
|
| 340 |
+
raise SystemExit(main())
|
tooluse1/parse.py
ADDED
|
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Extract tool calls from raw model output.
|
| 2 |
+
|
| 3 |
+
Models emit calls in many shapes: Qwen-native XML, OpenAI JSON, prose-wrapped
|
| 4 |
+
JSON in a code fence, plain-text `TOOL name | k=v`, or a bare argument object.
|
| 5 |
+
Extractors are tried in priority order and the first that yields calls wins.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
import json
|
| 11 |
+
import re
|
| 12 |
+
from typing import Any, Callable, Iterable, Mapping, Sequence
|
| 13 |
+
|
| 14 |
+
from .schema import ParseResult, ToolCall
|
| 15 |
+
from .tools import REGISTRY, canonical_name
|
| 16 |
+
|
| 17 |
+
_THINK_BLOCK = re.compile(r"<(think|thinking|reasoning|analysis)>.*?</\1>", re.S | re.I)
|
| 18 |
+
_THINK_OPEN = re.compile(r"<\|?(think|thinking|reasoning|analysis)\|?>.*?(?=</?tool_call|<tool_call|TOOL\s|\Z)", re.S | re.I)
|
| 19 |
+
_THINK_END_MARKER = "<｜end▁of▁thinking｜>"
|
| 20 |
+
_FENCE = re.compile(r"```(?:json|python|xml|yaml|tool_code)?\s*(.*?)```", re.S | re.I)
|
| 21 |
+
_XML_CALL = re.compile(r"<tool_call>(.*?)</tool_call>", re.S | re.I)
|
| 22 |
+
_XML_FUNCTION = re.compile(r"<function\s*=\s*([\w.\-]+)\s*>(.*?)(?:</function>|(?=<function\s*=)|(?=</tool_call>)|$)", re.S | re.I)
|
| 23 |
+
_XML_PARAM = re.compile(r"<parameter\s*=\s*([\w.\-]+)\s*>(.*?)(?:</parameter>|(?=<parameter\s*=)|$)", re.S | re.I)
|
| 24 |
+
_PLAIN_CALL = re.compile(r"(?:^|\n)\s*(?:\*\*)?TOOL(?:\s*CALL)?(?:\*\*)?\s*[:\-]?\s*([A-Za-z_][\w.\-]*)\s*(\|[^\n]*)?", re.I)
|
| 25 |
+
_PLAIN_CALL_ANY = re.compile(r"(?:\*\*)?TOOL(?:\s*CALL)?(?:\*\*)?\s*[:\-]?\s*[A-Za-z_][\w.\-]*\s*(?:\|[^\n]*)?", re.I)
|
| 26 |
+
_JSON_BLOB = re.compile(r"\{(?:[^{}]|\{[^{}]*\})*\}", re.S)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def strip_thinking(text: str) -> str:
|
| 30 |
+
if _THINK_END_MARKER in text:
|
| 31 |
+
text = text.rsplit(_THINK_END_MARKER, 1)[-1]
|
| 32 |
+
text = _THINK_BLOCK.sub(" ", text)
|
| 33 |
+
text = _THINK_OPEN.sub(" ", text)
|
| 34 |
+
return re.sub(r"</?(think|thinking|reasoning|analysis)>", " ", text, flags=re.I)
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def _coerce(value: Any) -> Any:
|
| 38 |
+
if not isinstance(value, str):
|
| 39 |
+
return value
|
| 40 |
+
text = value.strip()
|
| 41 |
+
if not text:
|
| 42 |
+
return ""
|
| 43 |
+
stripped = text.strip("`")
|
| 44 |
+
if stripped in ("true", "false"):
|
| 45 |
+
return stripped == "true"
|
| 46 |
+
if stripped in ("null", "none"):
|
| 47 |
+
return None
|
| 48 |
+
try:
|
| 49 |
+
return json.loads(stripped)
|
| 50 |
+
except (ValueError, TypeError):
|
| 51 |
+
pass
|
| 52 |
+
if re.fullmatch(r"[-+]?\d+", stripped):
|
| 53 |
+
return int(stripped)
|
| 54 |
+
if re.fullmatch(r"[-+]?\d*\.\d+([eE][-+]?\d+)?", stripped):
|
| 55 |
+
return float(stripped)
|
| 56 |
+
return text
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def _pair_args(raw: Mapping[str, Any] | str | None) -> dict[str, Any]:
|
| 60 |
+
if raw is None:
|
| 61 |
+
return {}
|
| 62 |
+
if isinstance(raw, str):
|
| 63 |
+
try:
|
| 64 |
+
raw = json.loads(raw)
|
| 65 |
+
except ValueError:
|
| 66 |
+
return {}
|
| 67 |
+
if not isinstance(raw, Mapping):
|
| 68 |
+
return {}
|
| 69 |
+
return {str(key): _coerce(value) for key, value in raw.items()}
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def _call_from_object(obj: Any) -> list[ToolCall]:
|
| 73 |
+
calls: list[ToolCall] = []
|
| 74 |
+
if isinstance(obj, list):
|
| 75 |
+
for item in obj:
|
| 76 |
+
calls.extend(_call_from_object(item))
|
| 77 |
+
return calls
|
| 78 |
+
if not isinstance(obj, Mapping):
|
| 79 |
+
return calls
|
| 80 |
+
|
| 81 |
+
if isinstance(obj.get("tool_calls"), list):
|
| 82 |
+
for item in obj["tool_calls"]:
|
| 83 |
+
calls.extend(_call_from_object(item))
|
| 84 |
+
return calls
|
| 85 |
+
|
| 86 |
+
function = obj.get("function")
|
| 87 |
+
if isinstance(function, Mapping):
|
| 88 |
+
name = function.get("name")
|
| 89 |
+
if isinstance(name, str):
|
| 90 |
+
args = _pair_args(function.get("arguments"))
|
| 91 |
+
if not args and isinstance(function.get("parameters"), (Mapping, str)):
|
| 92 |
+
args = _pair_args(function["parameters"])
|
| 93 |
+
return [ToolCall(name=canonical_name(name), args=args)]
|
| 94 |
+
if isinstance(function, str) and obj.get("arguments") is not None:
|
| 95 |
+
return [ToolCall(name=canonical_name(function), args=_pair_args(obj.get("arguments")))]
|
| 96 |
+
|
| 97 |
+
name = obj.get("name") or obj.get("tool") or obj.get("tool_name") or obj.get("function_name")
|
| 98 |
+
if isinstance(name, str) and canonical_name(name) in REGISTRY:
|
| 99 |
+
args_source = None
|
| 100 |
+
for key in ("arguments", "args", "parameters", "input", "parameter", "params"):
|
| 101 |
+
if key in obj:
|
| 102 |
+
args_source = obj[key]
|
| 103 |
+
break
|
| 104 |
+
args = _pair_args(args_source) if args_source is not None else {
|
| 105 |
+
key: _coerce(value)
|
| 106 |
+
for key, value in obj.items()
|
| 107 |
+
if key not in ("name", "tool", "tool_name", "function_name", "arguments", "args", "parameters", "input", "parameter", "params", "id", "type")
|
| 108 |
+
}
|
| 109 |
+
return [ToolCall(name=canonical_name(name), args=args)]
|
| 110 |
+
return calls
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def _extract_json(text: str) -> list[ToolCall]:
|
| 114 |
+
calls: list[ToolCall] = []
|
| 115 |
+
for match in _JSON_BLOB.finditer(text):
|
| 116 |
+
try:
|
| 117 |
+
obj = json.loads(match.group(0))
|
| 118 |
+
except ValueError:
|
| 119 |
+
continue
|
| 120 |
+
calls.extend(_call_from_object(obj))
|
| 121 |
+
return calls
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def _extract_xml(text: str) -> list[ToolCall]:
|
| 125 |
+
calls: list[ToolCall] = []
|
| 126 |
+
for block in _XML_CALL.findall(text):
|
| 127 |
+
inner = block.strip()
|
| 128 |
+
if inner.startswith("{"):
|
| 129 |
+
try:
|
| 130 |
+
calls.extend(_call_from_object(json.loads(inner)))
|
| 131 |
+
continue
|
| 132 |
+
except ValueError:
|
| 133 |
+
pass
|
| 134 |
+
for name, body in _XML_FUNCTION.findall(inner or text):
|
| 135 |
+
args = {key: _coerce(value) for key, value in _XML_PARAM.findall(body)}
|
| 136 |
+
calls.append(ToolCall(name=canonical_name(name), args=args))
|
| 137 |
+
if not calls and "<function" in text.lower():
|
| 138 |
+
for name, body in _XML_FUNCTION.findall(text):
|
| 139 |
+
args = {key: _coerce(value) for key, value in _XML_PARAM.findall(body)}
|
| 140 |
+
calls.append(ToolCall(name=canonical_name(name), args=args))
|
| 141 |
+
return calls
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def _extract_plain(text: str) -> list[ToolCall]:
|
| 145 |
+
calls: list[ToolCall] = []
|
| 146 |
+
for name, tail in _PLAIN_CALL.findall(text):
|
| 147 |
+
args: dict[str, Any] = {}
|
| 148 |
+
for chunk in (tail or "").split("|")[1:]:
|
| 149 |
+
if "=" in chunk:
|
| 150 |
+
key, value = chunk.split("=", 1)
|
| 151 |
+
args[key.strip().strip("`\"'")] = _coerce(value.strip().strip("`\"'"))
|
| 152 |
+
elif chunk.strip():
|
| 153 |
+
args.setdefault("value", _coerce(chunk.strip()))
|
| 154 |
+
calls.append(ToolCall(name=canonical_name(name), args=args))
|
| 155 |
+
return calls
|
| 156 |
+
|
| 157 |
+
|
| 158 |
+
def _extract_bare_json_args(text: str) -> list[ToolCall]:
|
| 159 |
+
"""A model may answer with arguments only, e.g. {"expression": "45+12"}."""
|
| 160 |
+
out: list[ToolCall] = []
|
| 161 |
+
for match in _JSON_BLOB.finditer(text):
|
| 162 |
+
try:
|
| 163 |
+
obj = json.loads(match.group(0))
|
| 164 |
+
except ValueError:
|
| 165 |
+
continue
|
| 166 |
+
if not isinstance(obj, Mapping) or not obj:
|
| 167 |
+
continue
|
| 168 |
+
keys = {str(key) for key in obj}
|
| 169 |
+
matches = [
|
| 170 |
+
spec.name
|
| 171 |
+
for spec in REGISTRY.values()
|
| 172 |
+
if set(spec.required) <= keys and keys <= set(spec.parameters)
|
| 173 |
+
]
|
| 174 |
+
if len(matches) == 1:
|
| 175 |
+
out.append(ToolCall(name=matches[0], args={str(k): _coerce(v) for k, v in obj.items()}))
|
| 176 |
+
return out
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
_STRATEGIES: tuple[tuple[str, Callable[[str], list[ToolCall]]], ...] = (
|
| 180 |
+
("xml", _extract_xml),
|
| 181 |
+
("json", _extract_json),
|
| 182 |
+
("plain", _extract_plain),
|
| 183 |
+
("bare_args", _extract_bare_json_args),
|
| 184 |
+
)
|
| 185 |
+
|
| 186 |
+
|
| 187 |
+
def parse_output(raw: str) -> ParseResult:
|
| 188 |
+
text = strip_thinking(raw or "")
|
| 189 |
+
searched: Sequence[str] = (text, *[m.group(1) for m in _FENCE.finditer(text)])
|
| 190 |
+
for source, extractor in _STRATEGIES:
|
| 191 |
+
for candidate in searched:
|
| 192 |
+
found = extractor(candidate)
|
| 193 |
+
if found:
|
| 194 |
+
return ParseResult(calls=tuple(found), source=source)
|
| 195 |
+
markers = ("<tool_call", "tool name", "to", "<function=", '"name"', '"tool_calls"', "||")
|
| 196 |
+
looks_like_attempt = "<tool_call" in text.lower() or bool(re.search(r"(?m)^\s*\|?(tool|to)\b", text, re.I)) or "<function=" in text.lower()
|
| 197 |
+
malformed = bool(looks_like_attempt) and not any(marker in text for marker in markers[:1])
|
| 198 |
+
return ParseResult(calls=(), source="none", malformed=malformed)
|
| 199 |
+
|
| 200 |
+
|
| 201 |
+
def answer_text(raw: str) -> str:
|
| 202 |
+
"""Model output with thinking blocks and call payloads removed."""
|
| 203 |
+
text = strip_thinking(raw or "")
|
| 204 |
+
text = _XML_CALL.sub(" ", text)
|
| 205 |
+
text = re.sub(r"<function\s*=.*?(?:</function>|$)", " ", text, flags=re.S | re.I)
|
| 206 |
+
text = _PLAIN_CALL_ANY.sub(" ", text)
|
| 207 |
+
text = _FENCE.sub(lambda m: m.group(1) if not m.group(1).strip().startswith("{") else " ", text)
|
| 208 |
+
for match in list(_JSON_BLOB.finditer(text)):
|
| 209 |
+
try:
|
| 210 |
+
obj = json.loads(match.group(0))
|
| 211 |
+
except ValueError:
|
| 212 |
+
continue
|
| 213 |
+
if isinstance(obj, Mapping) and ({"name", "tool_calls", "function"} & set(obj)) :
|
| 214 |
+
text = text[: match.start()] + " " + text[match.end() :]
|
| 215 |
+
return re.sub(r"\s+", " ", text).strip()
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
def iter_call_signatures(calls: Iterable[ToolCall]) -> list[str]:
|
| 219 |
+
return [call.canonical() for call in calls]
|
tooluse1/prompts.py
ADDED
|
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""System prompts and message rendering per calling protocol.
|
| 2 |
+
|
| 3 |
+
The protocol is a run-time axis, not a property of the data: the same case is
|
| 4 |
+
rendered in JSON, XML and plain-text form, so per-protocol scores are directly
|
| 5 |
+
comparable.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
from typing import Any, Mapping, Sequence
|
| 11 |
+
|
| 12 |
+
from .schema import Case, Message, Protocol
|
| 13 |
+
from .tools import ToolSpec, render_tools
|
| 14 |
+
|
| 15 |
+
IDENTITY_EN = (
|
| 16 |
+
"You are a helpful assistant developed by Hardcore AI. You answer in the same "
|
| 17 |
+
"language the user writes in. You are concise and precise."
|
| 18 |
+
)
|
| 19 |
+
|
| 20 |
+
IDENTITY_RU = (
|
| 21 |
+
"Ты полезный ассистент, разработанный Hardcore AI. Отвечай на том же языке, "
|
| 22 |
+
"на котором пишет пользователь. Будь кратким и точным."
|
| 23 |
+
)
|
| 24 |
+
|
| 25 |
+
_TOOL_HEADER = {
|
| 26 |
+
"en": "You have access to the following tools:",
|
| 27 |
+
"ru": "Тебе доступны следующие инструменты:",
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
_NO_TOOL_HEADER = {
|
| 31 |
+
"en": "You have no tools available for this conversation.",
|
| 32 |
+
"ru": "В этом диалоге у тебя нет доступных инструментов.",
|
| 33 |
+
}
|
| 34 |
+
|
| 35 |
+
_PROTOCOL_RULES = {
|
| 36 |
+
"json": {
|
| 37 |
+
"en": (
|
| 38 |
+
"To call a tool, reply with a single JSON object and nothing else:\n"
|
| 39 |
+
'{"name": "tool_name", "arguments": {"parameter": "value"}}\n'
|
| 40 |
+
"Use only the parameter names declared for that tool. Do not wrap the JSON in "
|
| 41 |
+
"markdown fences and do not add any prose around it."
|
| 42 |
+
),
|
| 43 |
+
"ru": (
|
| 44 |
+
"Чтобы вызвать инструмент, ответь одним JSON-объектом и больше ничем:\n"
|
| 45 |
+
'{"name": "tool_name", "arguments": {"parameter": "value"}}\n'
|
| 46 |
+
"Используй только те имена параметров, которые объявлены для этого инструмента. "
|
| 47 |
+
"Не оборачивай JSON в markdown и не добавляй текст вокруг него."
|
| 48 |
+
),
|
| 49 |
+
},
|
| 50 |
+
"xml": {
|
| 51 |
+
"en": (
|
| 52 |
+
"To call a tool, reply with a single XML call block and nothing else:\n"
|
| 53 |
+
"<tool_call>\n<function=tool_name>\n<parameter=parameter_name>\nvalue\n"
|
| 54 |
+
"</parameter>\n</function>\n</tool_call>\n"
|
| 55 |
+
"Use only the parameter names declared for that tool."
|
| 56 |
+
),
|
| 57 |
+
"ru": (
|
| 58 |
+
"Чтобы вызвать инструмент, ответь одним XML-блоком вызова и больше ничем:\n"
|
| 59 |
+
"<tool_call>\n<function=tool_name>\n<parameter=parameter_name>\nvalue\n"
|
| 60 |
+
"</parameter>\n</function>\n</tool_call>\n"
|
| 61 |
+
"Используй только те имена параметров, которые объявлены для этого инструмента."
|
| 62 |
+
),
|
| 63 |
+
},
|
| 64 |
+
"plain": {
|
| 65 |
+
"en": (
|
| 66 |
+
"To call a tool, reply with a single line in this exact format and nothing else:\n"
|
| 67 |
+
"TOOL tool_name | parameter=value | parameter2=value2\n"
|
| 68 |
+
"Use one `|` separated `name=value` segment per parameter, in the order declared "
|
| 69 |
+
"for that tool. Use only the parameter names declared for that tool."
|
| 70 |
+
),
|
| 71 |
+
"ru": (
|
| 72 |
+
"Чтобы вызвать инструмент, ответь одной строкой строго такого формата и больше ничем:\n"
|
| 73 |
+
"TOOL tool_name | parameter=value | parameter2=value2\n"
|
| 74 |
+
"Используй один сегмент `name=value`, разделённый `|`, на каждый параметр, в порядке "
|
| 75 |
+
"объявления инструмента. Используй только объявленные имена параметров."
|
| 76 |
+
),
|
| 77 |
+
},
|
| 78 |
+
}
|
| 79 |
+
|
| 80 |
+
_NO_CALL_RULE = {
|
| 81 |
+
"en": (
|
| 82 |
+
"If no tool applies to the user's request, do not emit a call. Answer the user "
|
| 83 |
+
"directly from your own knowledge in plain prose."
|
| 84 |
+
),
|
| 85 |
+
"ru": (
|
| 86 |
+
"Если ни один инструмент не подходит к запросу, не вызывай его. Ответь "
|
| 87 |
+
"пользователю напрямую обычным текстом, используя свои знания."
|
| 88 |
+
),
|
| 89 |
+
}
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def system_prompt(tools: Sequence[ToolSpec], protocol: Protocol, lang: str = "en") -> str:
|
| 93 |
+
language = lang if lang in ("en", "ru") else "en"
|
| 94 |
+
parts: list[str] = [IDENTITY_EN if language == "en" else IDENTITY_RU]
|
| 95 |
+
|
| 96 |
+
if tools:
|
| 97 |
+
parts.append(_TOOL_HEADER[language])
|
| 98 |
+
parts.append(render_tools(tools, protocol))
|
| 99 |
+
parts.append(_PROTOCOL_RULES[protocol][language])
|
| 100 |
+
else:
|
| 101 |
+
parts.append(_NO_TOOL_HEADER[language])
|
| 102 |
+
|
| 103 |
+
parts.append(_NO_CALL_RULE[language])
|
| 104 |
+
return "\n\n".join(parts)
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
def render_message(message: Message, protocol: Protocol) -> dict[str, str]:
|
| 108 |
+
if message.role == "assistant" and message.tool_name:
|
| 109 |
+
return {"role": "assistant", "content": render_call(message.tool_name, message.tool_args or {}, protocol)}
|
| 110 |
+
if message.role == "tool":
|
| 111 |
+
return {"role": "user", "content": f"Tool result for {message.tool_name}: {message.content}"}
|
| 112 |
+
return {"role": message.role, "content": message.content}
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def render_call(name: str, args: Mapping[str, Any], protocol: Protocol) -> str:
|
| 116 |
+
if protocol == "json":
|
| 117 |
+
import json
|
| 118 |
+
|
| 119 |
+
return json.dumps({"name": name, "arguments": dict(args)}, ensure_ascii=False)
|
| 120 |
+
if protocol == "xml":
|
| 121 |
+
body = "\n".join(f"<parameter={key}>\n{value}\n</parameter>" for key, value in args.items())
|
| 122 |
+
return f"<tool_call>\n<function={name}>\n{body}\n</function>\n</tool_call>"
|
| 123 |
+
segments = " | ".join(f"{key}={value}" for key, value in args.items())
|
| 124 |
+
return f"TOOL {name} | {segments}" if segments else f"TOOL {name}"
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def render_case(case: Case, protocol: Protocol, specs: Sequence[ToolSpec]) -> list[dict[str, str]]:
|
| 128 |
+
messages: list[dict[str, str]] = [
|
| 129 |
+
{"role": "system", "content": system_prompt(specs, protocol, case.lang)}
|
| 130 |
+
]
|
| 131 |
+
messages.extend(render_message(message, protocol) for message in case.history)
|
| 132 |
+
messages.append({"role": "user", "content": case.user})
|
| 133 |
+
return messages
|
tooluse1/report.py
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Markdown and JSON reporting for evaluation runs."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import json
|
| 6 |
+
from pathlib import Path
|
| 7 |
+
from typing import Any, Mapping, Sequence
|
| 8 |
+
|
| 9 |
+
from .grade import summarize
|
| 10 |
+
from .schema import CaseResult
|
| 11 |
+
|
| 12 |
+
_METRIC_LABELS = (
|
| 13 |
+
("score", "Overall"),
|
| 14 |
+
("correct_call_when_expected", "Correct call when a tool applies"),
|
| 15 |
+
("correct_abstention", "Correct abstention when no tool applies"),
|
| 16 |
+
("spurious_call_rate", "Spurious-call rate (lower is better)"),
|
| 17 |
+
("name_accuracy", "Tool selected correctly"),
|
| 18 |
+
("args_accuracy", "Arguments exactly correct"),
|
| 19 |
+
("answer_constraint_accuracy", "Answer constraints satisfied"),
|
| 20 |
+
("malformed_rate", "Malformed output rate"),
|
| 21 |
+
)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def percent(value: float) -> str:
|
| 25 |
+
return f"{value * 100:.1f}%"
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def _metric_rows(summary: Mapping[str, Any]) -> list[str]:
|
| 29 |
+
rows = ["| Metric | Score |", "|---|---|"]
|
| 30 |
+
for key, label in _METRIC_LABELS:
|
| 31 |
+
if key in summary:
|
| 32 |
+
rows.append(f"| {label} | {percent(summary[key])} |")
|
| 33 |
+
interval = summary.get("score_ci95")
|
| 34 |
+
if interval:
|
| 35 |
+
rows.append(f"| Overall 95% CI (Wilson) | {percent(interval[0])} – {percent(interval[1])} |")
|
| 36 |
+
return rows
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _breakdown(title: str, groups: Mapping[str, Mapping[str, Any]]) -> list[str]:
|
| 40 |
+
rows = [f"### {title}", "", "| Group | Cases | Score | Correct call | Abstention | Spurious calls |", "|---|---|---|---|---|---|"]
|
| 41 |
+
for name, group in groups.items():
|
| 42 |
+
rows.append(
|
| 43 |
+
f"| {name} | {group['cases']} | {percent(group['score'])} | "
|
| 44 |
+
f"{percent(group['correct_call_when_expected'])} | {percent(group['correct_abstention'])} | "
|
| 45 |
+
f"{percent(group['spurious_call_rate'])} |"
|
| 46 |
+
)
|
| 47 |
+
rows.append("")
|
| 48 |
+
return rows
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def render_markdown(summary: Mapping[str, Any], report: Mapping[str, Any] | None = None) -> str:
|
| 52 |
+
overall = summary["overall"]
|
| 53 |
+
lines = [
|
| 54 |
+
f"# ToolUse.1 — {summary.get('model') or 'unnamed model'}",
|
| 55 |
+
"",
|
| 56 |
+
f"- Protocol: **{summary.get('protocol')}**",
|
| 57 |
+
f"- Cases scored: **{summary.get('n_cases')}**",
|
| 58 |
+
"",
|
| 59 |
+
"## Summary",
|
| 60 |
+
"",
|
| 61 |
+
*_metric_rows(overall),
|
| 62 |
+
"",
|
| 63 |
+
]
|
| 64 |
+
for key, title in (
|
| 65 |
+
("by_protocol", "By protocol"),
|
| 66 |
+
("by_language", "By language"),
|
| 67 |
+
("by_category", "By category"),
|
| 68 |
+
("by_difficulty", "By difficulty"),
|
| 69 |
+
):
|
| 70 |
+
groups = summary.get(key) or {}
|
| 71 |
+
if groups:
|
| 72 |
+
lines.extend(_breakdown(title, groups))
|
| 73 |
+
if report is not None:
|
| 74 |
+
lines.extend(_failure_section(report))
|
| 75 |
+
return "\n".join(lines).rstrip() + "\n"
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
def _failure_section(report: Mapping[str, Any], limit: int = 40) -> list[str]:
|
| 79 |
+
failures = report.get("failures") or []
|
| 80 |
+
if not failures:
|
| 81 |
+
return []
|
| 82 |
+
lines = ["### Failing cases", "", "| Case | Protocol | Expected | Predicted |", "|---|---|---|---|"]
|
| 83 |
+
for item in failures[:limit]:
|
| 84 |
+
predicted = str(item.get("detail") or item.get("output") or "").replace("|", "\\|").replace("\n", " ")
|
| 85 |
+
lines.append(
|
| 86 |
+
f"| {item.get('case_id')} | {item.get('protocol')} | "
|
| 87 |
+
f"{'call' if item.get('expected_call') else 'no call'} | {predicted[:180]} |"
|
| 88 |
+
)
|
| 89 |
+
if len(failures) > limit:
|
| 90 |
+
lines.append("")
|
| 91 |
+
lines.append(f"_{len(failures) - limit} further failures omitted; see the raw results file._")
|
| 92 |
+
lines.append("")
|
| 93 |
+
return lines
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
def build_report(results: Sequence[CaseResult], model_id: str = "", protocol: str | None = None) -> dict[str, Any]:
|
| 97 |
+
summary = summarize(results, model_id=model_id, protocol=protocol) # type: ignore[arg-type]
|
| 98 |
+
failures = [result.to_dict() for result in results if not result.passed]
|
| 99 |
+
return {"summary": summary, "results": [result.to_dict() for result in results], "failures": failures}
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def write_report(report: Mapping[str, Any], directory: Path, stem: str) -> dict[str, Path]:
|
| 103 |
+
directory.mkdir(parents=True, exist_ok=True)
|
| 104 |
+
json_path = directory / f"{stem}.json"
|
| 105 |
+
markdown_path = directory / f"{stem}.md"
|
| 106 |
+
json_path.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
|
| 107 |
+
markdown_path.write_text(render_markdown(report["summary"], report), encoding="utf-8")
|
| 108 |
+
return {"json": json_path, "markdown": markdown_path}
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
def leaderboard_row(report: Mapping[str, Any], model_id: str, link: str = "", protocol: str = "all", notes: str = "") -> dict[str, Any]:
|
| 112 |
+
overall = report["summary"]["overall"]
|
| 113 |
+
return {
|
| 114 |
+
"model": model_id,
|
| 115 |
+
"link": link,
|
| 116 |
+
"protocol": protocol,
|
| 117 |
+
"score": overall["score"],
|
| 118 |
+
"score_ci95": overall["score_ci95"],
|
| 119 |
+
"correct_call": overall["correct_call_when_expected"],
|
| 120 |
+
"abstention": overall["correct_abstention"],
|
| 121 |
+
"spurious_call_rate": overall["spurious_call_rate"],
|
| 122 |
+
"n_cases": report["summary"]["n_cases"],
|
| 123 |
+
"notes": notes,
|
| 124 |
+
}
|
tooluse1/runners.py
ADDED
|
@@ -0,0 +1,396 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Model backends.
|
| 2 |
+
|
| 3 |
+
Every backend implements `complete(prompt) -> str`. The harness treats a raised
|
| 4 |
+
exception as a failed case rather than aborting the run, so a single flaky
|
| 5 |
+
response cannot destroy a multi-hour evaluation.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
import json
|
| 11 |
+
import urllib.error
|
| 12 |
+
import urllib.request
|
| 13 |
+
from dataclasses import dataclass
|
| 14 |
+
from typing import Any, Mapping, Sequence
|
| 15 |
+
|
| 16 |
+
from .prompts import render_call
|
| 17 |
+
from .schema import Case, ExpectCall, ExpectNoCall, Protocol
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
@dataclass(frozen=True)
|
| 21 |
+
class Prompt:
|
| 22 |
+
case: Case
|
| 23 |
+
protocol: Protocol
|
| 24 |
+
messages: tuple[Mapping[str, str], ...]
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class Runner:
|
| 28 |
+
name: str = "runner"
|
| 29 |
+
|
| 30 |
+
def complete(self, prompt: Prompt) -> str: # pragma: no cover - interface
|
| 31 |
+
raise NotImplementedError
|
| 32 |
+
|
| 33 |
+
def complete_many(self, prompts: Sequence[Prompt]) -> list[str]:
|
| 34 |
+
"""Default: sequential. Backends that can batch override this."""
|
| 35 |
+
return [self.complete(prompt) for prompt in prompts]
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
@dataclass(frozen=True)
|
| 39 |
+
class BatchOutput:
|
| 40 |
+
text: str | None = None
|
| 41 |
+
error: str = ""
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def run_batch(runner: Runner, prompts: Sequence[Prompt], batch_size: int) -> list[BatchOutput]:
|
| 45 |
+
"""Run a chunk, degrading to one-by-one if the backend refuses the batch."""
|
| 46 |
+
if batch_size > 1:
|
| 47 |
+
try:
|
| 48 |
+
return [BatchOutput(text=text) for text in runner.complete_many(prompts)]
|
| 49 |
+
except Exception: # noqa: BLE001 - a poisoned batch must not lose its neighbours
|
| 50 |
+
pass
|
| 51 |
+
outputs: list[BatchOutput] = []
|
| 52 |
+
for prompt in prompts:
|
| 53 |
+
try:
|
| 54 |
+
outputs.append(BatchOutput(text=runner.complete(prompt)))
|
| 55 |
+
except Exception as error: # noqa: BLE001
|
| 56 |
+
outputs.append(BatchOutput(error=f"runner error: {type(error).__name__}: {error}"))
|
| 57 |
+
return outputs
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
class OracleRunner(Runner):
|
| 61 |
+
"""Emits exactly what the corpus expects. Used to validate the grader itself."""
|
| 62 |
+
|
| 63 |
+
name = "oracle"
|
| 64 |
+
|
| 65 |
+
def complete(self, prompt: Prompt) -> str:
|
| 66 |
+
expect = prompt.case.expect
|
| 67 |
+
if isinstance(expect, ExpectCall):
|
| 68 |
+
return render_call(expect.name, expect.args, prompt.protocol)
|
| 69 |
+
needles: list[str] = list(expect.answer_contains_all)
|
| 70 |
+
needles.extend(group[0] for group in expect.answer_contains_any if group)
|
| 71 |
+
if needles:
|
| 72 |
+
return "Here is the answer: " + ", ".join(needles) + "."
|
| 73 |
+
return "I can answer that from what I already know."
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
class NullRunner(Runner):
|
| 77 |
+
"""Answers politely and never calls a tool."""
|
| 78 |
+
|
| 79 |
+
name = "null"
|
| 80 |
+
|
| 81 |
+
def complete(self, prompt: Prompt) -> str:
|
| 82 |
+
return "I am not sure about that."
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
class FencedRunner(OracleRunner):
|
| 86 |
+
"""Wraps every call in a markdown fence behind a sentence of prose."""
|
| 87 |
+
|
| 88 |
+
name = "fenced"
|
| 89 |
+
|
| 90 |
+
def complete(self, prompt: Prompt) -> str:
|
| 91 |
+
raw = super().complete(prompt)
|
| 92 |
+
if isinstance(prompt.case.expect, ExpectCall):
|
| 93 |
+
fence = "json" if prompt.protocol == "json" else "text"
|
| 94 |
+
return f"Sure, calling the tool now.\n\n```{fence}\n{raw}\n```\n"
|
| 95 |
+
return raw
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
class ThinkingRunner(OracleRunner):
|
| 99 |
+
"""Prefixes every call with a reasoning block, the way a thinking model does."""
|
| 100 |
+
|
| 101 |
+
name = "thinking"
|
| 102 |
+
|
| 103 |
+
def complete(self, prompt: Prompt) -> str:
|
| 104 |
+
raw = super().complete(prompt)
|
| 105 |
+
if isinstance(prompt.case.expect, ExpectCall):
|
| 106 |
+
return f"<thinking>\nThe user wants a tool call. Pick {prompt.case.expect.name}.\n</thinking>\n{raw}"
|
| 107 |
+
return f"<thinking>No tool applies here.</thinking>\n{raw}"
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
class ScriptedRunner(Runner):
|
| 111 |
+
"""Deterministic per-case overrides, for unit tests and demos."""
|
| 112 |
+
|
| 113 |
+
name = "scripted"
|
| 114 |
+
|
| 115 |
+
def __init__(self, overrides: Mapping[str, str], default: str = ""):
|
| 116 |
+
self._overrides = dict(overrides)
|
| 117 |
+
self._default = default
|
| 118 |
+
|
| 119 |
+
def complete(self, prompt: Prompt) -> str:
|
| 120 |
+
return self._overrides.get(prompt.case.id, self._default)
|
| 121 |
+
|
| 122 |
+
|
| 123 |
+
class OpenAICompatibleRunner(Runner):
|
| 124 |
+
"""Any `/chat/completions` endpoint: vLLM, llama.cpp server, Ollama, OpenAI, OpenRouter."""
|
| 125 |
+
|
| 126 |
+
def __init__(
|
| 127 |
+
self,
|
| 128 |
+
model: str,
|
| 129 |
+
base_url: str = "https://api.openai.com/v1",
|
| 130 |
+
api_key: str = "",
|
| 131 |
+
max_tokens: int = 256,
|
| 132 |
+
temperature: float = 0.0,
|
| 133 |
+
timeout: float = 120.0,
|
| 134 |
+
name: str | None = None,
|
| 135 |
+
):
|
| 136 |
+
self.model = model
|
| 137 |
+
self.base_url = base_url.rstrip("/")
|
| 138 |
+
self.api_key = api_key
|
| 139 |
+
self.max_tokens = max_tokens
|
| 140 |
+
self.temperature = temperature
|
| 141 |
+
self.timeout = timeout
|
| 142 |
+
self.name = name or model
|
| 143 |
+
|
| 144 |
+
def complete(self, prompt: Prompt) -> str:
|
| 145 |
+
body = json.dumps(
|
| 146 |
+
{
|
| 147 |
+
"model": self.model,
|
| 148 |
+
"messages": [dict(message) for message in prompt.messages],
|
| 149 |
+
"max_tokens": self.max_tokens,
|
| 150 |
+
"temperature": self.temperature,
|
| 151 |
+
"stream": False,
|
| 152 |
+
}
|
| 153 |
+
).encode("utf-8")
|
| 154 |
+
request = urllib.request.Request(
|
| 155 |
+
f"{self.base_url}/chat/completions",
|
| 156 |
+
data=body,
|
| 157 |
+
method="POST",
|
| 158 |
+
headers={
|
| 159 |
+
"Content-Type": "application/json",
|
| 160 |
+
**({"Authorization": f"Bearer {self.api_key}"} if self.api_key else {}),
|
| 161 |
+
},
|
| 162 |
+
)
|
| 163 |
+
try:
|
| 164 |
+
with urllib.request.urlopen(request, timeout=self.timeout) as response:
|
| 165 |
+
payload = json.loads(response.read().decode("utf-8"))
|
| 166 |
+
except urllib.error.HTTPError as error:
|
| 167 |
+
detail = error.read().decode("utf-8", "replace")[:400]
|
| 168 |
+
raise RuntimeError(f"HTTP {error.code}: {detail}") from error
|
| 169 |
+
choices = payload.get("choices") or []
|
| 170 |
+
if not choices:
|
| 171 |
+
return ""
|
| 172 |
+
message = choices[0].get("message") or {}
|
| 173 |
+
content = message.get("content")
|
| 174 |
+
if isinstance(content, list):
|
| 175 |
+
return "".join(part.get("text", "") for part in content if isinstance(part, Mapping))
|
| 176 |
+
return content or message.get("reasoning_content") or ""
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
class TransformersRunner(Runner):
|
| 180 |
+
"""Local HF checkpoint, greedy decoding.
|
| 181 |
+
|
| 182 |
+
Handles three realities of modern checkpoints: vision-capable conditional
|
| 183 |
+
generation architectures (`AutoModelForImageTextToText`), the `dtype` rename,
|
| 184 |
+
and chat templates with a reasoning switch that must be turned off for tool
|
| 185 |
+
calls to be emitted instead of a thinking trace.
|
| 186 |
+
"""
|
| 187 |
+
|
| 188 |
+
def __init__(
|
| 189 |
+
self,
|
| 190 |
+
model_id: str,
|
| 191 |
+
max_new_tokens: int = 256,
|
| 192 |
+
dtype: str = "auto",
|
| 193 |
+
device_map: str = "auto",
|
| 194 |
+
trust_remote_code: bool = True,
|
| 195 |
+
revision: str | None = None,
|
| 196 |
+
system_override: str | None = None,
|
| 197 |
+
disable_thinking: bool = True,
|
| 198 |
+
):
|
| 199 |
+
import torch
|
| 200 |
+
|
| 201 |
+
self._torch = torch
|
| 202 |
+
self.name = model_id
|
| 203 |
+
self.model_id = model_id
|
| 204 |
+
self.max_new_tokens = max_new_tokens
|
| 205 |
+
self.system_override = system_override
|
| 206 |
+
self.disable_thinking = disable_thinking
|
| 207 |
+
self.tokenizer = _load_processor(model_id, trust_remote_code, revision)
|
| 208 |
+
if getattr(self.tokenizer, "pad_token", None) is None:
|
| 209 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
| 210 |
+
self.model, self.load_info = _load_model(model_id, dtype, device_map, trust_remote_code, revision)
|
| 211 |
+
self.model.eval()
|
| 212 |
+
|
| 213 |
+
def _render(self, messages: Sequence[Mapping[str, str]]) -> str:
|
| 214 |
+
kwargs: dict[str, Any] = {"tokenize": False, "add_generation_prompt": True}
|
| 215 |
+
if self.disable_thinking:
|
| 216 |
+
kwargs["enable_thinking"] = False
|
| 217 |
+
try:
|
| 218 |
+
return self.tokenizer.apply_chat_template(messages, **kwargs)
|
| 219 |
+
except TypeError:
|
| 220 |
+
kwargs.pop("enable_thinking", None)
|
| 221 |
+
return self.tokenizer.apply_chat_template(messages, **kwargs)
|
| 222 |
+
|
| 223 |
+
def _prepare(self, prompt: Prompt) -> str:
|
| 224 |
+
messages = [dict(message) for message in prompt.messages]
|
| 225 |
+
if self.system_override and messages and messages[0]["role"] == "system":
|
| 226 |
+
messages[0]["content"] = self.system_override
|
| 227 |
+
return self._render(messages)
|
| 228 |
+
|
| 229 |
+
def _inputs(self, texts: Sequence[str], padding: bool) -> dict[str, Any]:
|
| 230 |
+
previous_side = getattr(self.tokenizer, "padding_side", "right")
|
| 231 |
+
if padding:
|
| 232 |
+
# Decoder-only generation must continue after the padded prefix, so the
|
| 233 |
+
# padding belongs on the left; the slice below then trims uniformly.
|
| 234 |
+
self.tokenizer.padding_side = "left"
|
| 235 |
+
try:
|
| 236 |
+
try:
|
| 237 |
+
inputs = self.tokenizer(text=list(texts), return_tensors="pt", padding=padding)
|
| 238 |
+
except TypeError:
|
| 239 |
+
inputs = self.tokenizer(list(texts), return_tensors="pt", padding=padding)
|
| 240 |
+
finally:
|
| 241 |
+
self.tokenizer.padding_side = previous_side
|
| 242 |
+
if hasattr(inputs, "to"):
|
| 243 |
+
inputs = inputs.to(self.model.device)
|
| 244 |
+
return {key: value for key, value in dict(inputs).items()}
|
| 245 |
+
|
| 246 |
+
def _generate(self, inputs: dict[str, Any]) -> list[str]:
|
| 247 |
+
input_length = inputs["input_ids"].shape[-1]
|
| 248 |
+
with self._torch.inference_mode():
|
| 249 |
+
generated = self.model.generate(
|
| 250 |
+
**inputs,
|
| 251 |
+
max_new_tokens=self.max_new_tokens,
|
| 252 |
+
do_sample=False,
|
| 253 |
+
pad_token_id=self.tokenizer.pad_token_id or self.tokenizer.eos_token_id,
|
| 254 |
+
)
|
| 255 |
+
return [self.tokenizer.decode(row[input_length:], skip_special_tokens=True) for row in generated]
|
| 256 |
+
|
| 257 |
+
def complete(self, prompt: Prompt) -> str:
|
| 258 |
+
return self._generate(self._inputs([self._prepare(prompt)], padding=False))[0]
|
| 259 |
+
|
| 260 |
+
def complete_many(self, prompts: Sequence[Prompt]) -> list[str]:
|
| 261 |
+
"""One padded `generate` call per batch.
|
| 262 |
+
|
| 263 |
+
Memory is not the constraint here — a 0.8B model in bfloat16 leaves several
|
| 264 |
+
gigabytes free on a 6 GB card — GPU *utilisation* is. Batching turns a decode
|
| 265 |
+
loop that idles between kernel launches into a saturated one.
|
| 266 |
+
"""
|
| 267 |
+
if not prompts:
|
| 268 |
+
return []
|
| 269 |
+
if len(prompts) == 1:
|
| 270 |
+
return self.complete(prompts[0])
|
| 271 |
+
return self._generate(self._inputs([self._prepare(prompt) for prompt in prompts], padding=True))
|
| 272 |
+
|
| 273 |
+
|
| 274 |
+
def _load_processor(model_id: str, trust_remote_code: bool, revision: str | None):
|
| 275 |
+
from transformers import AutoProcessor, AutoTokenizer
|
| 276 |
+
|
| 277 |
+
errors: list[str] = []
|
| 278 |
+
for loader in (AutoTokenizer, AutoProcessor):
|
| 279 |
+
try:
|
| 280 |
+
return loader.from_pretrained(model_id, trust_remote_code=trust_remote_code, revision=revision)
|
| 281 |
+
except Exception as error: # noqa: BLE001
|
| 282 |
+
errors.append(f"{loader.__name__}: {type(error).__name__}: {error}")
|
| 283 |
+
raise RuntimeError("could not load a tokenizer/processor for " + model_id + "\n" + "\n".join(errors))
|
| 284 |
+
|
| 285 |
+
|
| 286 |
+
LOAD_CANDIDATES: tuple[str, ...] = ("AutoModelForCausalLM", "AutoModelForImageTextToText", "AutoModel")
|
| 287 |
+
|
| 288 |
+
_IGNORABLE_MISSING: tuple[str, ...] = (
|
| 289 |
+
"inv_freq",
|
| 290 |
+
"rotary",
|
| 291 |
+
"causal_mask",
|
| 292 |
+
"position_ids",
|
| 293 |
+
"mtp.",
|
| 294 |
+
"logit_scale",
|
| 295 |
+
"temperature",
|
| 296 |
+
)
|
| 297 |
+
|
| 298 |
+
|
| 299 |
+
def _core_missing(missing: Sequence[str]) -> list[str]:
|
| 300 |
+
return [key for key in missing if not any(token in key for token in _IGNORABLE_MISSING)]
|
| 301 |
+
|
| 302 |
+
|
| 303 |
+
def _load_model(
|
| 304 |
+
model_id: str,
|
| 305 |
+
dtype: str,
|
| 306 |
+
device_map: str,
|
| 307 |
+
trust_remote_code: bool,
|
| 308 |
+
revision: str | None,
|
| 309 |
+
):
|
| 310 |
+
"""Load the first architecture whose weights actually match the checkpoint.
|
| 311 |
+
|
| 312 |
+
`from_pretrained` emits a warning, not an error, when a class expects keys the
|
| 313 |
+
checkpoint does not contain — the result is a randomly initialised model that still
|
| 314 |
+
produces fluent-looking text. Nova ships text-only keys (`model.layers.*`) and also
|
| 315 |
+
carries a vision tower, so the vision wrapper loads "successfully" and generates
|
| 316 |
+
noise. Every candidate is therefore verified against its loading report before it is
|
| 317 |
+
accepted, and the caller is told how it was loaded.
|
| 318 |
+
"""
|
| 319 |
+
import transformers
|
| 320 |
+
|
| 321 |
+
errors: list[str] = []
|
| 322 |
+
for name in LOAD_CANDIDATES:
|
| 323 |
+
loader = getattr(transformers, name, None)
|
| 324 |
+
if loader is None:
|
| 325 |
+
continue
|
| 326 |
+
for dtype_kwarg in ("dtype", "torch_dtype"):
|
| 327 |
+
try:
|
| 328 |
+
model, info = loader.from_pretrained(
|
| 329 |
+
model_id,
|
| 330 |
+
trust_remote_code=trust_remote_code,
|
| 331 |
+
revision=revision,
|
| 332 |
+
device_map=device_map,
|
| 333 |
+
output_loading_info=True,
|
| 334 |
+
**{dtype_kwarg: dtype},
|
| 335 |
+
)
|
| 336 |
+
except Exception as error: # noqa: BLE001
|
| 337 |
+
errors.append(f"{name}.{dtype_kwarg}: {type(error).__name__}: {error}")
|
| 338 |
+
continue
|
| 339 |
+
missing = _core_missing(list(info.get("missing_keys") or []))
|
| 340 |
+
if missing:
|
| 341 |
+
errors.append(
|
| 342 |
+
f"{name}: {len(missing)} checkpoint weights missing, e.g. {missing[0]} — refusing a random model"
|
| 343 |
+
)
|
| 344 |
+
del model
|
| 345 |
+
continue
|
| 346 |
+
unexpected = len(list(info.get("unexpected_keys") or []))
|
| 347 |
+
note = f"{name} ({dtype_kwarg}={dtype})" + (f", {unexpected} extra keys ignored" if unexpected else "")
|
| 348 |
+
return model, note
|
| 349 |
+
raise RuntimeError("could not load " + model_id + ":\n" + "\n".join(errors))
|
| 350 |
+
|
| 351 |
+
|
| 352 |
+
class ScriptedFileRunner(ScriptedRunner):
|
| 353 |
+
"""Loads `{"case_id": "output"}` from a JSON file — hand-graded runs, ablations."""
|
| 354 |
+
|
| 355 |
+
@staticmethod
|
| 356 |
+
def from_file(path: str, default: str = "") -> "ScriptedFileRunner":
|
| 357 |
+
with open(path, encoding="utf-8") as handle:
|
| 358 |
+
payload: dict[str, Any] = json.load(handle)
|
| 359 |
+
return ScriptedFileRunner({str(key): str(value) for key, value in payload.items()}, default=default)
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
def build_runner(backend: str, **kwargs: Any) -> Runner:
|
| 363 |
+
if backend == "oracle":
|
| 364 |
+
return OracleRunner()
|
| 365 |
+
if backend == "null":
|
| 366 |
+
return NullRunner()
|
| 367 |
+
if backend == "fenced":
|
| 368 |
+
return FencedRunner()
|
| 369 |
+
if backend == "thinking":
|
| 370 |
+
return ThinkingRunner()
|
| 371 |
+
if backend in ("transformers", "local"):
|
| 372 |
+
return TransformersRunner(
|
| 373 |
+
model_id=kwargs["model"],
|
| 374 |
+
max_new_tokens=kwargs.get("max_new_tokens", 256),
|
| 375 |
+
dtype=kwargs.get("dtype", "auto"),
|
| 376 |
+
revision=kwargs.get("revision"),
|
| 377 |
+
disable_thinking=kwargs.get("disable_thinking", True),
|
| 378 |
+
)
|
| 379 |
+
if backend in ("openai", "vllm", "ollama", "llamacpp"):
|
| 380 |
+
return OpenAICompatibleRunner(
|
| 381 |
+
model=kwargs["model"],
|
| 382 |
+
base_url=kwargs.get("base_url", "http://localhost:8000/v1"),
|
| 383 |
+
api_key=kwargs.get("api_key", ""),
|
| 384 |
+
max_tokens=kwargs.get("max_new_tokens", 256),
|
| 385 |
+
)
|
| 386 |
+
if backend == "scripted":
|
| 387 |
+
return ScriptedFileRunner.from_file(kwargs["scripted_file"])
|
| 388 |
+
raise ValueError(f"unknown backend: {backend}")
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def rendered_calls(expect: Any, protocol: Protocol) -> Sequence[str]:
|
| 392 |
+
if isinstance(expect, ExpectCall):
|
| 393 |
+
return (render_call(expect.name, expect.args, protocol),)
|
| 394 |
+
if isinstance(expect, ExpectNoCall):
|
| 395 |
+
return ()
|
| 396 |
+
return ()
|
tooluse1/schema.py
ADDED
|
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Core data types for ToolUse.1.
|
| 2 |
+
|
| 3 |
+
Errors are values: parsers return `Result` objects instead of raising, so the
|
| 4 |
+
harness can report *why* a model failed rather than crashing mid-run.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
from __future__ import annotations
|
| 8 |
+
|
| 9 |
+
import json
|
| 10 |
+
from dataclasses import dataclass, field
|
| 11 |
+
from typing import Any, Literal, Mapping, Sequence
|
| 12 |
+
|
| 13 |
+
Role = Literal["system", "user", "assistant", "tool"]
|
| 14 |
+
|
| 15 |
+
Lang = Literal["en", "ru"]
|
| 16 |
+
|
| 17 |
+
Protocol = Literal["json", "xml", "plain"]
|
| 18 |
+
|
| 19 |
+
PROTOCOLS: tuple[Protocol, ...] = ("json", "xml", "plain")
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
@dataclass(frozen=True)
|
| 23 |
+
class Message:
|
| 24 |
+
role: Role
|
| 25 |
+
content: str
|
| 26 |
+
tool_name: str | None = None
|
| 27 |
+
tool_args: Mapping[str, Any] | None = None
|
| 28 |
+
|
| 29 |
+
def to_dict(self) -> dict[str, Any]:
|
| 30 |
+
out: dict[str, Any] = {"role": self.role, "content": self.content}
|
| 31 |
+
if self.tool_name is not None:
|
| 32 |
+
out["tool_name"] = self.tool_name
|
| 33 |
+
if self.tool_args is not None:
|
| 34 |
+
out["tool_args"] = dict(self.tool_args)
|
| 35 |
+
return out
|
| 36 |
+
|
| 37 |
+
@staticmethod
|
| 38 |
+
def from_dict(raw: Mapping[str, Any]) -> "Message":
|
| 39 |
+
return Message(
|
| 40 |
+
role=raw["role"],
|
| 41 |
+
content=raw.get("content", ""),
|
| 42 |
+
tool_name=raw.get("tool_name"),
|
| 43 |
+
tool_args=raw.get("tool_args"),
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
@dataclass(frozen=True)
|
| 48 |
+
class ExpectCall:
|
| 49 |
+
kind: Literal["call"] = "call"
|
| 50 |
+
name: str = ""
|
| 51 |
+
args: Mapping[str, Any] = field(default_factory=dict)
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
@dataclass(frozen=True)
|
| 55 |
+
class ExpectNoCall:
|
| 56 |
+
kind: Literal["no_call"] = "no_call"
|
| 57 |
+
answer_contains_any: Sequence[Sequence[str]] = ()
|
| 58 |
+
answer_contains_all: Sequence[str] = ()
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
Expectation = ExpectCall | ExpectNoCall
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
@dataclass(frozen=True)
|
| 65 |
+
class Case:
|
| 66 |
+
id: str
|
| 67 |
+
lang: Lang
|
| 68 |
+
category: str
|
| 69 |
+
difficulty: str
|
| 70 |
+
tools: tuple[str, ...]
|
| 71 |
+
user: str
|
| 72 |
+
expect: Expectation
|
| 73 |
+
history: tuple[Message, ...] = ()
|
| 74 |
+
notes: str = ""
|
| 75 |
+
|
| 76 |
+
def messages(self) -> tuple[Message, ...]:
|
| 77 |
+
return self.history + (Message(role="user", content=self.user),)
|
| 78 |
+
|
| 79 |
+
def to_dict(self) -> dict[str, Any]:
|
| 80 |
+
raw: dict[str, Any] = {
|
| 81 |
+
"id": self.id,
|
| 82 |
+
"lang": self.lang,
|
| 83 |
+
"category": self.category,
|
| 84 |
+
"difficulty": self.difficulty,
|
| 85 |
+
"tools": list(self.tools),
|
| 86 |
+
"user": self.user,
|
| 87 |
+
"expect": _expect_to_dict(self.expect),
|
| 88 |
+
}
|
| 89 |
+
if self.history:
|
| 90 |
+
raw["history"] = [m.to_dict() for m in self.history]
|
| 91 |
+
if self.notes:
|
| 92 |
+
raw["notes"] = self.notes
|
| 93 |
+
return raw
|
| 94 |
+
|
| 95 |
+
def to_json(self) -> str:
|
| 96 |
+
return json.dumps(self.to_dict(), ensure_ascii=False, sort_keys=True)
|
| 97 |
+
|
| 98 |
+
@staticmethod
|
| 99 |
+
def from_dict(raw: Mapping[str, Any]) -> "Case":
|
| 100 |
+
return Case(
|
| 101 |
+
id=raw["id"],
|
| 102 |
+
lang=raw["lang"],
|
| 103 |
+
category=raw["category"],
|
| 104 |
+
difficulty=raw["difficulty"],
|
| 105 |
+
tools=tuple(raw["tools"]),
|
| 106 |
+
user=raw["user"],
|
| 107 |
+
expect=_expect_from_dict(raw["expect"]),
|
| 108 |
+
history=tuple(Message.from_dict(m) for m in raw.get("history", ())),
|
| 109 |
+
notes=raw.get("notes", ""),
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def _expect_to_dict(expect: Expectation) -> dict[str, Any]:
|
| 114 |
+
if expect.kind == "call":
|
| 115 |
+
return {"kind": "call", "name": expect.name, "args": dict(expect.args)}
|
| 116 |
+
out: dict[str, Any] = {"kind": "no_call"}
|
| 117 |
+
if expect.answer_contains_any:
|
| 118 |
+
out["answer_contains_any"] = [list(group) for group in expect.answer_contains_any]
|
| 119 |
+
if expect.answer_contains_all:
|
| 120 |
+
out["answer_contains_all"] = list(expect.answer_contains_all)
|
| 121 |
+
return out
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def _expect_from_dict(raw: Mapping[str, Any]) -> Expectation:
|
| 125 |
+
if raw["kind"] == "call":
|
| 126 |
+
return ExpectCall(name=raw["name"], args=raw.get("args", {}))
|
| 127 |
+
return ExpectNoCall(
|
| 128 |
+
answer_contains_any=tuple(tuple(g) for g in raw.get("answer_contains_any", ())),
|
| 129 |
+
answer_contains_all=tuple(raw.get("answer_contains_all", ())),
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
@dataclass(frozen=True)
|
| 134 |
+
class ToolCall:
|
| 135 |
+
name: str
|
| 136 |
+
args: Mapping[str, Any]
|
| 137 |
+
|
| 138 |
+
def canonical(self) -> str:
|
| 139 |
+
return json.dumps({"name": self.name, "args": dict(self.args)}, ensure_ascii=False, sort_keys=True)
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
@dataclass(frozen=True)
|
| 143 |
+
class ParseResult:
|
| 144 |
+
calls: tuple[ToolCall, ...]
|
| 145 |
+
source: str
|
| 146 |
+
malformed: bool = False
|
| 147 |
+
detail: str = ""
|
| 148 |
+
|
| 149 |
+
@property
|
| 150 |
+
def call(self) -> ToolCall | None:
|
| 151 |
+
return self.calls[0] if self.calls else None
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
@dataclass(frozen=True)
|
| 155 |
+
class CaseResult:
|
| 156 |
+
case_id: str
|
| 157 |
+
protocol: Protocol
|
| 158 |
+
lang: Lang
|
| 159 |
+
category: str
|
| 160 |
+
difficulty: str
|
| 161 |
+
passed: bool
|
| 162 |
+
called: bool
|
| 163 |
+
expected_call: bool
|
| 164 |
+
name_correct: bool
|
| 165 |
+
args_exact: bool
|
| 166 |
+
answer_ok: bool | None
|
| 167 |
+
malformed: bool
|
| 168 |
+
output: str
|
| 169 |
+
detail: str = ""
|
| 170 |
+
|
| 171 |
+
def to_dict(self) -> dict[str, Any]:
|
| 172 |
+
return {
|
| 173 |
+
"case_id": self.case_id,
|
| 174 |
+
"protocol": self.protocol,
|
| 175 |
+
"lang": self.lang,
|
| 176 |
+
"category": self.category,
|
| 177 |
+
"difficulty": self.difficulty,
|
| 178 |
+
"passed": self.passed,
|
| 179 |
+
"called": self.called,
|
| 180 |
+
"expected_call": self.expected_call,
|
| 181 |
+
"name_correct": self.name_correct,
|
| 182 |
+
"args_exact": self.args_exact,
|
| 183 |
+
"answer_ok": self.answer_ok,
|
| 184 |
+
"malformed": self.malformed,
|
| 185 |
+
"output": self.output,
|
| 186 |
+
"detail": self.detail,
|
| 187 |
+
}
|
tooluse1/tools.py
ADDED
|
@@ -0,0 +1,615 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Tool registry: schemas, protocol renderers, and argument normalisation.
|
| 2 |
+
|
| 3 |
+
Grading rests on normalisation, not string equality. Every tool declares how its
|
| 4 |
+
arguments are canonicalised so that semantically identical calls compare equal
|
| 5 |
+
(`Berlin` / `berlin ` / `Берлин`, `300` / `300.0`, `45+12` / `57`).
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
import ast
|
| 11 |
+
import json
|
| 12 |
+
import math
|
| 13 |
+
import re
|
| 14 |
+
import unicodedata
|
| 15 |
+
from dataclasses import dataclass
|
| 16 |
+
from typing import Any, Callable, Mapping, Sequence
|
| 17 |
+
|
| 18 |
+
from .schema import Protocol
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
@dataclass(frozen=True)
|
| 22 |
+
class ToolSpec:
|
| 23 |
+
name: str
|
| 24 |
+
description: str
|
| 25 |
+
parameters: Mapping[str, str]
|
| 26 |
+
required: tuple[str, ...]
|
| 27 |
+
normalizers: Mapping[str, Callable[[Any], Any]]
|
| 28 |
+
optional: tuple[str, ...] = ()
|
| 29 |
+
|
| 30 |
+
def schema(self) -> dict[str, Any]:
|
| 31 |
+
return {
|
| 32 |
+
"name": self.name,
|
| 33 |
+
"description": self.description,
|
| 34 |
+
"parameters": {
|
| 35 |
+
"type": "object",
|
| 36 |
+
"properties": {
|
| 37 |
+
key: {"type": param_type, "description": self.parameters[key]}
|
| 38 |
+
for key, param_type in self.parameters.items()
|
| 39 |
+
},
|
| 40 |
+
"required": list(self.required),
|
| 41 |
+
},
|
| 42 |
+
}
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
_QUOTES = str.maketrans({"«": "", "»": "", "“": "", "”": "", '"': "", "'": ""})
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def _squash(value: Any) -> str:
|
| 49 |
+
text = unicodedata.normalize("NFKC", str(value))
|
| 50 |
+
text = text.translate(_QUOTES).replace("\u0451", "\u0435").replace("\u00eb", "e")
|
| 51 |
+
text = re.sub(r"\s+", " ", text).strip().lower()
|
| 52 |
+
return text
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
_CITY_ALIASES = {
|
| 56 |
+
"spb": "saint petersburg",
|
| 57 |
+
"st petersburg": "saint petersburg",
|
| 58 |
+
"st. petersburg": "saint petersburg",
|
| 59 |
+
"sankt peterburg": "saint petersburg",
|
| 60 |
+
"petersburg": "saint petersburg",
|
| 61 |
+
"\u043f\u0438\u0442\u0435\u0440": "saint petersburg",
|
| 62 |
+
"\u0441\u0430\u043d\u043a\u0442-\u043f\u0435\u0442\u0435\u0440\u0431\u0443\u0440\u0433": "saint petersburg",
|
| 63 |
+
"\u0441\u043f\u0431": "saint petersburg",
|
| 64 |
+
"nyc": "new york",
|
| 65 |
+
"new york city": "new york",
|
| 66 |
+
"\u043d\u044c\u044e-\u0439\u043e\u0440\u043a": "new york",
|
| 67 |
+
"msk": "moscow",
|
| 68 |
+
"\u043c\u043e\u0441\u043a\u0432\u0430": "moscow",
|
| 69 |
+
"kiev": "kyiv",
|
| 70 |
+
"\u043a\u0438\u0435\u0432": "kyiv",
|
| 71 |
+
"\u043a\u0438\u0457\u0432": "kyiv",
|
| 72 |
+
"\u0431\u0435\u0440\u043b\u0438\u043d": "berlin",
|
| 73 |
+
"\u043b\u043e\u043d\u0434\u043e\u043d": "london",
|
| 74 |
+
"\u0442\u043e\u043a\u0438\u043e": "tokyo",
|
| 75 |
+
"\u043f\u0430\u0440\u0438\u0436": "paris",
|
| 76 |
+
"\u0434\u0443\u0431\u0430\u0439": "dubai",
|
| 77 |
+
"\u0430\u043b\u043c\u0430\u0442\u044b": "almaty",
|
| 78 |
+
"\u043a\u0430\u0437\u0430\u043d\u044c": "kazan",
|
| 79 |
+
}
|
| 80 |
+
|
| 81 |
+
_LANGUAGE_ALIASES = {
|
| 82 |
+
"english": "en",
|
| 83 |
+
"eng": "en",
|
| 84 |
+
"en-us": "en",
|
| 85 |
+
"en-gb": "en",
|
| 86 |
+
"\u0430\u043d\u0433\u043b\u0438\u0439\u0441\u043a\u0438\u0439": "en",
|
| 87 |
+
"russian": "ru",
|
| 88 |
+
"rus": "ru",
|
| 89 |
+
"ru-ru": "ru",
|
| 90 |
+
"\u0440\u0443\u0441\u0441\u043a\u0438\u0439": "ru",
|
| 91 |
+
"german": "de",
|
| 92 |
+
"\u043d\u0435\u043c\u0435\u0446\u043a\u0438\u0439": "de",
|
| 93 |
+
"germany": "de",
|
| 94 |
+
"spanish": "es",
|
| 95 |
+
"\u0438\u0441\u043f\u0430\u043d\u0441\u043a\u0438\u0439": "es",
|
| 96 |
+
"french": "fr",
|
| 97 |
+
"\u0444\u0440\u0430\u043d\u0446\u0443\u0437\u0441\u043a\u0438\u0439": "fr",
|
| 98 |
+
"italian": "it",
|
| 99 |
+
"\u0438\u0442\u0430\u043b\u044c\u044f\u043d\u0441\u043a\u0438\u0439": "it",
|
| 100 |
+
"japanese": "ja",
|
| 101 |
+
"\u044f\u043f\u043e\u043d\u0441\u043a\u0438\u0439": "ja",
|
| 102 |
+
"chinese": "zh",
|
| 103 |
+
"\u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0439": "zh",
|
| 104 |
+
"portuguese": "pt",
|
| 105 |
+
"\u043f\u043e\u0440\u0442\u0443\u0433\u0430\u043b\u044c\u0441\u043a\u0438\u0439": "pt",
|
| 106 |
+
"turkish": "tr",
|
| 107 |
+
"\u0442\u0443\u0440\u0435\u0446\u043a\u0438\u0439": "tr",
|
| 108 |
+
"kazakh": "kk",
|
| 109 |
+
"\u043a\u0430\u0437\u0430\u0445\u0441\u043a\u0438\u0439": "kk",
|
| 110 |
+
"ukrainian": "uk",
|
| 111 |
+
"\u0443\u043a\u0440\u0430\u0438\u043d\u0441\u043a\u0438\u0439": "uk",
|
| 112 |
+
"arabic": "ar",
|
| 113 |
+
"\u0430\u0440\u0430\u0431\u0441\u043a\u0438\u0439": "ar",
|
| 114 |
+
"korean": "ko",
|
| 115 |
+
"\u043a\u043e\u0440\u0435\u0439\u0441\u043a\u0438\u0439": "ko",
|
| 116 |
+
"dutch": "nl",
|
| 117 |
+
"\u0433\u043e\u043b\u043b\u0430\u043d\u0434\u0441\u043a\u0438\u0439": "nl",
|
| 118 |
+
"polish": "pl",
|
| 119 |
+
"\u043f\u043e\u043b\u044c\u0441\u043a\u0438\u0439": "pl",
|
| 120 |
+
"hindi": "hi",
|
| 121 |
+
}
|
| 122 |
+
|
| 123 |
+
_CURRENCY_ALIASES = {
|
| 124 |
+
"$": "USD",
|
| 125 |
+
"usd": "USD",
|
| 126 |
+
"us dollar": "USD",
|
| 127 |
+
"us dollars": "USD",
|
| 128 |
+
"dollar": "USD",
|
| 129 |
+
"dollars": "USD",
|
| 130 |
+
"\u0434\u043e\u043b\u043b\u0430\u0440": "USD",
|
| 131 |
+
"\u0434\u043e\u043b\u043b\u0430\u0440\u043e\u0432": "USD",
|
| 132 |
+
"usd.": "USD",
|
| 133 |
+
"\u20ac": "EUR",
|
| 134 |
+
"eur": "EUR",
|
| 135 |
+
"euro": "EUR",
|
| 136 |
+
"euros": "EUR",
|
| 137 |
+
"\u0435\u0432\u0440\u043e": "EUR",
|
| 138 |
+
"\u20bd": "RUB",
|
| 139 |
+
"rub": "RUB",
|
| 140 |
+
"ruble": "RUB",
|
| 141 |
+
"rubles": "RUB",
|
| 142 |
+
"rouble": "RUB",
|
| 143 |
+
"\u0440\u0443\u0431": "RUB",
|
| 144 |
+
"\u0440\u0443\u0431\u043b\u044c": "RUB",
|
| 145 |
+
"\u0440\u0443\u0431\u043b\u0435\u0439": "RUB",
|
| 146 |
+
"\u20b8": "KZT",
|
| 147 |
+
"kzt": "KZT",
|
| 148 |
+
"tenge": "KZT",
|
| 149 |
+
"\u0442\u0435\u043d\u0433\u0435": "KZT",
|
| 150 |
+
"\u00a3": "GBP",
|
| 151 |
+
"gbp": "GBP",
|
| 152 |
+
"pound": "GBP",
|
| 153 |
+
"pounds": "GBP",
|
| 154 |
+
"sterling": "GBP",
|
| 155 |
+
"\u0444\u0443\u043d\u0442": "GBP",
|
| 156 |
+
"\u00a5": "JPY",
|
| 157 |
+
"jpy": "JPY",
|
| 158 |
+
"yen": "JPY",
|
| 159 |
+
"\u0438\u0435\u043d\u0430": "JPY",
|
| 160 |
+
"\u00a5\u00a5": "CNY",
|
| 161 |
+
"cny": "CNY",
|
| 162 |
+
"rmb": "CNY",
|
| 163 |
+
"yuan": "CNY",
|
| 164 |
+
"\u044e\u0430\u043d\u044c": "CNY",
|
| 165 |
+
"try": "TRY",
|
| 166 |
+
"\u20ba": "TRY",
|
| 167 |
+
"lira": "TRY",
|
| 168 |
+
"\u043b\u0438\u0440\u0430": "TRY",
|
| 169 |
+
"\u20b4": "UAH",
|
| 170 |
+
"uah": "UAH",
|
| 171 |
+
"hryvnia": "UAH",
|
| 172 |
+
"\u0433\u0440\u0438\u0432\u043d\u0430": "UAH",
|
| 173 |
+
"\u20b9": "INR",
|
| 174 |
+
"inr": "INR",
|
| 175 |
+
"rupee": "INR",
|
| 176 |
+
"rupees": "INR",
|
| 177 |
+
"chf": "CHF",
|
| 178 |
+
"franc": "CHF",
|
| 179 |
+
"cad": "CAD",
|
| 180 |
+
"aud": "AUD",
|
| 181 |
+
"brl": "BRL",
|
| 182 |
+
"real": "BRL",
|
| 183 |
+
"sek": "SEK",
|
| 184 |
+
"nok": "NOK",
|
| 185 |
+
"pln": "PLN",
|
| 186 |
+
"zloty": "PLN",
|
| 187 |
+
"czk": "CZK",
|
| 188 |
+
"krone": "NOK",
|
| 189 |
+
}
|
| 190 |
+
|
| 191 |
+
_UNIT_ALIASES = {
|
| 192 |
+
"km": "km",
|
| 193 |
+
"kilometer": "km",
|
| 194 |
+
"kilometers": "km",
|
| 195 |
+
"kilometre": "km",
|
| 196 |
+
"kilometres": "km",
|
| 197 |
+
"\u043a\u043c": "km",
|
| 198 |
+
"\u043a\u0438\u043b\u043e\u043c\u0435\u0442\u0440": "km",
|
| 199 |
+
"\u043a\u0438\u043b\u043e\u043c\u0435\u0442\u0440\u043e\u0432": "km",
|
| 200 |
+
"mile": "mi",
|
| 201 |
+
"miles": "mi",
|
| 202 |
+
"mi": "mi",
|
| 203 |
+
"\u043c\u0438\u043b\u044f": "mi",
|
| 204 |
+
"\u043c\u0438\u043b\u044c": "mi",
|
| 205 |
+
"\u043c\u0438\u043b\u0438": "mi",
|
| 206 |
+
"m": "m",
|
| 207 |
+
"meter": "m",
|
| 208 |
+
"meters": "m",
|
| 209 |
+
"metre": "m",
|
| 210 |
+
"\u043c\u0435\u0442\u0440": "m",
|
| 211 |
+
"\u043c\u0435\u0442\u0440\u043e\u0432": "m",
|
| 212 |
+
"ft": "ft",
|
| 213 |
+
"foot": "ft",
|
| 214 |
+
"feet": "ft",
|
| 215 |
+
"\u0444\u0443\u0442": "ft",
|
| 216 |
+
"\u0444\u0443\u0442\u043e\u0432": "ft",
|
| 217 |
+
"cm": "cm",
|
| 218 |
+
"centimeter": "cm",
|
| 219 |
+
"centimeters": "cm",
|
| 220 |
+
"\u0441\u043c": "cm",
|
| 221 |
+
"inch": "in",
|
| 222 |
+
"inches": "in",
|
| 223 |
+
"in": "in",
|
| 224 |
+
"\u0434\u044e\u0439\u043c": "in",
|
| 225 |
+
"\u0434\u044e\u0439\u043c\u043e\u0432": "in",
|
| 226 |
+
"mm": "mm",
|
| 227 |
+
"millimeter": "mm",
|
| 228 |
+
"\u043c\u043c": "mm",
|
| 229 |
+
"kg": "kg",
|
| 230 |
+
"kilogram": "kg",
|
| 231 |
+
"kilograms": "kg",
|
| 232 |
+
"\u043a\u0433": "kg",
|
| 233 |
+
"kilogramm": "kg",
|
| 234 |
+
"lb": "lb",
|
| 235 |
+
"lbs": "lb",
|
| 236 |
+
"pound": "lb",
|
| 237 |
+
"pounds": "lb",
|
| 238 |
+
"\u0444\u0443\u043d\u0442": "lb",
|
| 239 |
+
"g": "g",
|
| 240 |
+
"gram": "g",
|
| 241 |
+
"grams": "g",
|
| 242 |
+
"\u0433": "g",
|
| 243 |
+
"\u0433\u0440\u0430\u043c\u043c": "g",
|
| 244 |
+
"\u0433\u0440\u0430\u043c\u043c\u043e\u0432": "g",
|
| 245 |
+
"c": "celsius",
|
| 246 |
+
"celsius": "celsius",
|
| 247 |
+
"\u00b0c": "celsius",
|
| 248 |
+
"\u0433\u0440\u0430\u0434\u0443\u0441\u043e\u0432": "celsius",
|
| 249 |
+
"\u0446\u0435\u043b\u044c\u0441\u0438\u0439": "celsius",
|
| 250 |
+
"f": "fahrenheit",
|
| 251 |
+
"fahrenheit": "fahrenheit",
|
| 252 |
+
"\u00b0f": "fahrenheit",
|
| 253 |
+
"farenheit": "fahrenheit",
|
| 254 |
+
"kph": "km/h",
|
| 255 |
+
"km/h": "km/h",
|
| 256 |
+
"kmh": "km/h",
|
| 257 |
+
"\u043a\u043c/\u0447": "km/h",
|
| 258 |
+
"mph": "mph",
|
| 259 |
+
"m/s": "m/s",
|
| 260 |
+
"l": "l",
|
| 261 |
+
"liter": "l",
|
| 262 |
+
"liters": "l",
|
| 263 |
+
"litre": "l",
|
| 264 |
+
"\u043b": "l",
|
| 265 |
+
"\u043b\u0438\u0442\u0440": "l",
|
| 266 |
+
"\u043b\u0438\u0442\u0440\u043e\u0432": "l",
|
| 267 |
+
"ml": "ml",
|
| 268 |
+
"\u043c\u043b": "ml",
|
| 269 |
+
"gal": "gal",
|
| 270 |
+
"gallon": "gal",
|
| 271 |
+
"gallons": "gal",
|
| 272 |
+
"hour": "h",
|
| 273 |
+
"hours": "h",
|
| 274 |
+
"h": "h",
|
| 275 |
+
"\u0447\u0430\u0441": "h",
|
| 276 |
+
"\u0447\u0430\u0441\u043e\u0432": "h",
|
| 277 |
+
"\u0447": "h",
|
| 278 |
+
"minute": "min",
|
| 279 |
+
"minutes": "min",
|
| 280 |
+
"\u043c\u0438\u043d\u0443\u0442": "min",
|
| 281 |
+
"\u043c\u0438\u043d": "min",
|
| 282 |
+
"day": "day",
|
| 283 |
+
"days": "day",
|
| 284 |
+
"\u0434\u0435\u043d\u044c": "day",
|
| 285 |
+
"\u0434\u043d\u0435\u0439": "day",
|
| 286 |
+
}
|
| 287 |
+
|
| 288 |
+
_TIMEZONE_ALIASES = {
|
| 289 |
+
"utc": "UTC",
|
| 290 |
+
"gmt": "UTC",
|
| 291 |
+
"moscow": "Europe/Moscow",
|
| 292 |
+
"msk": "Europe/Moscow",
|
| 293 |
+
"moscow time": "Europe/Moscow",
|
| 294 |
+
"\u043c\u043e\u0441\u043a\u0432\u0430": "Europe/Moscow",
|
| 295 |
+
"saint petersburg": "Europe/Moscow",
|
| 296 |
+
"\u043c\u0441\u043a": "Europe/Moscow",
|
| 297 |
+
"berlin": "Europe/Berlin",
|
| 298 |
+
"london": "Europe/London",
|
| 299 |
+
"paris": "Europe/Paris",
|
| 300 |
+
"new york": "America/New_York",
|
| 301 |
+
"nyc": "America/New_York",
|
| 302 |
+
"los angeles": "America/Los_Angeles",
|
| 303 |
+
"tokyo": "Asia/Tokyo",
|
| 304 |
+
"dubai": "Asia/Dubai",
|
| 305 |
+
"almaty": "Asia/Almaty",
|
| 306 |
+
"shanghai": "Asia/Shanghai",
|
| 307 |
+
"sydney": "Australia/Sydney",
|
| 308 |
+
"kyiv": "Europe/Kyiv",
|
| 309 |
+
"kiev": "Europe/Kyiv",
|
| 310 |
+
"kazan": "Europe/Moscow",
|
| 311 |
+
"warsaw": "Europe/Warsaw",
|
| 312 |
+
"istanbul": "Europe/Istanbul",
|
| 313 |
+
"singapore": "Asia/Singapore",
|
| 314 |
+
}
|
| 315 |
+
|
| 316 |
+
_MATH_FUNCS: dict[str, Callable[..., float]] = {
|
| 317 |
+
"sqrt": math.sqrt,
|
| 318 |
+
"abs": abs,
|
| 319 |
+
"round": round,
|
| 320 |
+
"floor": math.floor,
|
| 321 |
+
"ceil": math.ceil,
|
| 322 |
+
"min": min,
|
| 323 |
+
"max": max,
|
| 324 |
+
"pow": pow,
|
| 325 |
+
"sin": math.sin,
|
| 326 |
+
"cos": math.cos,
|
| 327 |
+
"tan": math.tan,
|
| 328 |
+
"log": math.log,
|
| 329 |
+
"log10": math.log10,
|
| 330 |
+
"exp": math.exp,
|
| 331 |
+
"pi": lambda: math.pi,
|
| 332 |
+
"e": lambda: math.e,
|
| 333 |
+
}
|
| 334 |
+
|
| 335 |
+
|
| 336 |
+
def eval_expression(expression: Any) -> float | None:
|
| 337 |
+
"""Evaluate arithmetic safely. Returns None when the expression is not arithmetic."""
|
| 338 |
+
if isinstance(expression, bool):
|
| 339 |
+
return None
|
| 340 |
+
if isinstance(expression, (int, float)):
|
| 341 |
+
return float(expression)
|
| 342 |
+
text = _squash(expression).replace("\u00d7", "*").replace("\u00f7", "/").replace("\u0445", "*")
|
| 343 |
+
text = text.replace("^", "**").replace(",", "").replace("%", "/100")
|
| 344 |
+
text = re.sub(r"(?<=\d)\s+(\d)", r"*\1", text)
|
| 345 |
+
if not text or not re.fullmatch(r"[0-9a-z_+\-*/().\s]*", text):
|
| 346 |
+
return None
|
| 347 |
+
try:
|
| 348 |
+
tree = ast.parse(text, mode="eval")
|
| 349 |
+
except SyntaxError:
|
| 350 |
+
return None
|
| 351 |
+
|
| 352 |
+
def walk(node: ast.AST) -> float:
|
| 353 |
+
if isinstance(node, ast.Expression):
|
| 354 |
+
return walk(node.body)
|
| 355 |
+
if isinstance(node, ast.Constant) and isinstance(node.value, (int, float)):
|
| 356 |
+
return float(node.value)
|
| 357 |
+
if isinstance(node, ast.UnaryOp) and isinstance(node.op, (ast.UAdd, ast.USub)):
|
| 358 |
+
value = walk(node.operand)
|
| 359 |
+
return value if isinstance(node.op, ast.UAdd) else -value
|
| 360 |
+
if isinstance(node, ast.BinOp):
|
| 361 |
+
left, right = walk(node.left), walk(node.right)
|
| 362 |
+
ops: dict[type, Callable[[float, float], float]] = {
|
| 363 |
+
ast.Add: lambda a, b: a + b,
|
| 364 |
+
ast.Sub: lambda a, b: a - b,
|
| 365 |
+
ast.Mult: lambda a, b: a * b,
|
| 366 |
+
ast.Div: lambda a, b: a / b,
|
| 367 |
+
ast.Pow: lambda a, b: a**b,
|
| 368 |
+
ast.Mod: lambda a, b: a % b,
|
| 369 |
+
ast.FloorDiv: lambda a, b: a // b,
|
| 370 |
+
}
|
| 371 |
+
for node_type, op in ops.items():
|
| 372 |
+
if isinstance(node.op, node_type):
|
| 373 |
+
return op(left, right)
|
| 374 |
+
raise ValueError("operator")
|
| 375 |
+
if isinstance(node, ast.Call):
|
| 376 |
+
if not isinstance(node.func, ast.Name):
|
| 377 |
+
raise ValueError("call")
|
| 378 |
+
func = _MATH_FUNCS.get(node.func.id)
|
| 379 |
+
if func is None or node.keywords:
|
| 380 |
+
raise ValueError("function")
|
| 381 |
+
return float(func(*[walk(arg) for arg in node.args]))
|
| 382 |
+
if isinstance(node, ast.Name):
|
| 383 |
+
func = _MATH_FUNCS.get(node.id)
|
| 384 |
+
if func is None:
|
| 385 |
+
raise ValueError("name")
|
| 386 |
+
return float(func())
|
| 387 |
+
raise ValueError("node")
|
| 388 |
+
|
| 389 |
+
try:
|
| 390 |
+
result = walk(tree)
|
| 391 |
+
except Exception:
|
| 392 |
+
return None
|
| 393 |
+
if not math.isfinite(result):
|
| 394 |
+
return None
|
| 395 |
+
return result
|
| 396 |
+
|
| 397 |
+
|
| 398 |
+
def norm_number(value: Any) -> float | None:
|
| 399 |
+
if isinstance(value, bool):
|
| 400 |
+
return None
|
| 401 |
+
if isinstance(value, (int, float)):
|
| 402 |
+
return float(value)
|
| 403 |
+
text = _squash(value).replace(",", ".").replace(" ", "")
|
| 404 |
+
text = re.sub(r"[^\d.\-e+]", "", text)
|
| 405 |
+
if not text:
|
| 406 |
+
return None
|
| 407 |
+
try:
|
| 408 |
+
return float(text)
|
| 409 |
+
except ValueError:
|
| 410 |
+
return None
|
| 411 |
+
|
| 412 |
+
|
| 413 |
+
def norm_expression(value: Any) -> Any:
|
| 414 |
+
evaluated = eval_expression(value)
|
| 415 |
+
if evaluated is not None:
|
| 416 |
+
return round(evaluated, 6)
|
| 417 |
+
return _squash(value).replace(" ", "")
|
| 418 |
+
|
| 419 |
+
|
| 420 |
+
def norm_from_aliases(aliases: Mapping[str, str], fallback: Callable[[str], str] | None = None) -> Callable[[Any], Any]:
|
| 421 |
+
table = {_squash(key): _squash(value) for key, value in aliases.items()}
|
| 422 |
+
tail: Callable[[str], str] = fallback or (lambda text: text)
|
| 423 |
+
|
| 424 |
+
def normalizer(value: Any) -> Any:
|
| 425 |
+
key = _squash(value)
|
| 426 |
+
key = re.sub(r"^(в|to|from|into|in)\s+", "", key).strip()
|
| 427 |
+
return table.get(key, tail(key))
|
| 428 |
+
|
| 429 |
+
return normalizer
|
| 430 |
+
|
| 431 |
+
|
| 432 |
+
_norm_city = norm_from_aliases(_CITY_ALIASES, lambda t: t.rstrip(".,!?"))
|
| 433 |
+
_norm_language = norm_from_aliases(_LANGUAGE_ALIASES)
|
| 434 |
+
_norm_currency = norm_from_aliases(_CURRENCY_ALIASES)
|
| 435 |
+
_norm_unit = norm_from_aliases(_UNIT_ALIASES)
|
| 436 |
+
_norm_timezone = norm_from_aliases(_TIMEZONE_ALIASES)
|
| 437 |
+
_norm_text = lambda value: re.sub(r"\s+", " ", _squash(value)).strip(" .,!?\"'")
|
| 438 |
+
|
| 439 |
+
|
| 440 |
+
def norm_query(value: Any) -> Any:
|
| 441 |
+
text = _norm_text(value)
|
| 442 |
+
text = re.sub(r"[^\w\s\u0400-\u04ff]", " ", text)
|
| 443 |
+
return re.sub(r"\s+", " ", text).strip()
|
| 444 |
+
|
| 445 |
+
|
| 446 |
+
def norm_seconds(value: Any) -> Any:
|
| 447 |
+
number = norm_number(value)
|
| 448 |
+
if number is None:
|
| 449 |
+
return _norm_text(value)
|
| 450 |
+
return int(round(number))
|
| 451 |
+
|
| 452 |
+
|
| 453 |
+
REGISTRY: dict[str, ToolSpec] = {
|
| 454 |
+
"calculator": ToolSpec(
|
| 455 |
+
name="calculator",
|
| 456 |
+
description="Evaluates a mathematical expression and returns the numeric result.",
|
| 457 |
+
parameters={
|
| 458 |
+
"expression": "string: the arithmetic expression to evaluate, e.g. '(45 + 12) * 3'"
|
| 459 |
+
},
|
| 460 |
+
required=("expression",),
|
| 461 |
+
normalizers={"expression": norm_expression},
|
| 462 |
+
),
|
| 463 |
+
"get_weather": ToolSpec(
|
| 464 |
+
name="get_weather",
|
| 465 |
+
description="Returns the current weather for a city.",
|
| 466 |
+
parameters={
|
| 467 |
+
"city": "string: city name",
|
| 468 |
+
"unit": "string, optional: 'celsius' or 'fahrenheit'",
|
| 469 |
+
},
|
| 470 |
+
required=("city",),
|
| 471 |
+
optional=("unit",),
|
| 472 |
+
normalizers={"city": _norm_city, "unit": _norm_unit},
|
| 473 |
+
),
|
| 474 |
+
"web_search": ToolSpec(
|
| 475 |
+
name="web_search",
|
| 476 |
+
description="Searches the web and returns ranked results for a query.",
|
| 477 |
+
parameters={"query": "string: search query"},
|
| 478 |
+
required=("query",),
|
| 479 |
+
normalizers={"query": norm_query},
|
| 480 |
+
),
|
| 481 |
+
"translate": ToolSpec(
|
| 482 |
+
name="translate",
|
| 483 |
+
description="Translates text into a target language.",
|
| 484 |
+
parameters={
|
| 485 |
+
"text": "string: the text to translate",
|
| 486 |
+
"target_language": "string: ISO 639-1 code or language name",
|
| 487 |
+
},
|
| 488 |
+
required=("text", "target_language"),
|
| 489 |
+
normalizers={"text": _norm_text, "target_language": _norm_language},
|
| 490 |
+
),
|
| 491 |
+
"convert_currency": ToolSpec(
|
| 492 |
+
name="convert_currency",
|
| 493 |
+
description="Converts an amount from one currency to another using current rates.",
|
| 494 |
+
parameters={
|
| 495 |
+
"amount": "number: amount to convert",
|
| 496 |
+
"from_currency": "string: ISO 4217 code or currency name",
|
| 497 |
+
"to_currency": "string: ISO 4217 code or currency name",
|
| 498 |
+
},
|
| 499 |
+
required=("amount", "from_currency", "to_currency"),
|
| 500 |
+
normalizers={
|
| 501 |
+
"amount": norm_number,
|
| 502 |
+
"from_currency": _norm_currency,
|
| 503 |
+
"to_currency": _norm_currency,
|
| 504 |
+
},
|
| 505 |
+
),
|
| 506 |
+
"convert_units": ToolSpec(
|
| 507 |
+
name="convert_units",
|
| 508 |
+
description="Converts a value between measurement units.",
|
| 509 |
+
parameters={
|
| 510 |
+
"value": "number: value to convert",
|
| 511 |
+
"from_unit": "string: source unit",
|
| 512 |
+
"to_unit": "string: target unit",
|
| 513 |
+
},
|
| 514 |
+
required=("value", "from_unit", "to_unit"),
|
| 515 |
+
normalizers={"value": norm_number, "from_unit": _norm_unit, "to_unit": _norm_unit},
|
| 516 |
+
),
|
| 517 |
+
"set_timer": ToolSpec(
|
| 518 |
+
name="set_timer",
|
| 519 |
+
description="Starts a countdown timer for the given duration in seconds.",
|
| 520 |
+
parameters={"seconds": "integer: duration of the timer in seconds"},
|
| 521 |
+
required=("seconds",),
|
| 522 |
+
normalizers={"seconds": norm_seconds},
|
| 523 |
+
),
|
| 524 |
+
"save_note": ToolSpec(
|
| 525 |
+
name="save_note",
|
| 526 |
+
description="Saves a note with a title and body text.",
|
| 527 |
+
parameters={"title": "string: short note title", "content": "string: note body"},
|
| 528 |
+
required=("title", "content"),
|
| 529 |
+
normalizers={"title": _norm_text, "content": _norm_text},
|
| 530 |
+
),
|
| 531 |
+
"get_time": ToolSpec(
|
| 532 |
+
name="get_time",
|
| 533 |
+
description="Returns the current date and time in a given timezone.",
|
| 534 |
+
parameters={"timezone": "string: IANA timezone name or city"},
|
| 535 |
+
required=("timezone",),
|
| 536 |
+
normalizers={"timezone": _norm_timezone},
|
| 537 |
+
),
|
| 538 |
+
}
|
| 539 |
+
|
| 540 |
+
TOOL_NAMES: tuple[str, ...] = tuple(REGISTRY)
|
| 541 |
+
|
| 542 |
+
# Handles seen in the wild that mean the same tool.
|
| 543 |
+
NAME_ALIASES: dict[str, str] = {
|
| 544 |
+
"calc": "calculator",
|
| 545 |
+
"calculate": "calculator",
|
| 546 |
+
"math": "calculator",
|
| 547 |
+
"evaluate_expression": "calculator",
|
| 548 |
+
"weather": "get_weather",
|
| 549 |
+
"weather_lookup": "get_weather",
|
| 550 |
+
"get_current_weather": "get_weather",
|
| 551 |
+
"search": "web_search",
|
| 552 |
+
"search_web": "web_search",
|
| 553 |
+
"google_search": "web_search",
|
| 554 |
+
"currency_converter": "convert_currency",
|
| 555 |
+
"convert": "convert_currency",
|
| 556 |
+
"exchange_rate": "convert_currency",
|
| 557 |
+
"unit_conversion": "convert_units",
|
| 558 |
+
"convert_measurement": "convert_units",
|
| 559 |
+
"timer": "set_timer",
|
| 560 |
+
"create_timer": "set_timer",
|
| 561 |
+
"start_timer": "set_timer",
|
| 562 |
+
"note": "save_note",
|
| 563 |
+
"add_note": "save_note",
|
| 564 |
+
"create_note": "save_note",
|
| 565 |
+
"time": "get_time",
|
| 566 |
+
"current_time": "get_time",
|
| 567 |
+
"get_current_time": "get_time",
|
| 568 |
+
"translation": "translate",
|
| 569 |
+
}
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
def canonical_name(name: Any) -> str:
|
| 573 |
+
key = _squash(name)
|
| 574 |
+
return NAME_ALIASES.get(key, key)
|
| 575 |
+
|
| 576 |
+
|
| 577 |
+
def tools_for(names: Sequence[str]) -> tuple[ToolSpec, ...]:
|
| 578 |
+
return tuple(REGISTRY[name] for name in names if name in REGISTRY)
|
| 579 |
+
|
| 580 |
+
|
| 581 |
+
def render_tools_json(specs: Sequence[ToolSpec]) -> str:
|
| 582 |
+
payload = {"tools": [spec.schema() for spec in specs]}
|
| 583 |
+
return json.dumps(payload, ensure_ascii=False, indent=2)
|
| 584 |
+
|
| 585 |
+
|
| 586 |
+
def render_tools_xml(specs: Sequence[ToolSpec]) -> str:
|
| 587 |
+
lines: list[str] = []
|
| 588 |
+
for spec in specs:
|
| 589 |
+
lines.append(f"<tool name=\"{spec.name}\">")
|
| 590 |
+
lines.append(f" <description>{spec.description}</description>")
|
| 591 |
+
lines.append(" <parameters>")
|
| 592 |
+
for key, param_type in spec.parameters.items():
|
| 593 |
+
marker = " required" if key in spec.required else ""
|
| 594 |
+
lines.append(f" <parameter name=\"{key}\" type=\"{param_type.split(':')[0]}\"{marker}/>")
|
| 595 |
+
lines.append(" </parameters>")
|
| 596 |
+
lines.append("</tool>")
|
| 597 |
+
return "\n".join(lines)
|
| 598 |
+
|
| 599 |
+
|
| 600 |
+
def render_tools_plain(specs: Sequence[ToolSpec]) -> str:
|
| 601 |
+
lines: list[str] = []
|
| 602 |
+
for spec in specs:
|
| 603 |
+
params = ", ".join(
|
| 604 |
+
f"{key}{'' if key in spec.required else ' (optional)'}" for key in spec.parameters
|
| 605 |
+
)
|
| 606 |
+
lines.append(f"- {spec.name}({params}): {spec.description}")
|
| 607 |
+
return "\n".join(lines)
|
| 608 |
+
|
| 609 |
+
|
| 610 |
+
def render_tools(specs: Sequence[ToolSpec], protocol: Protocol) -> str:
|
| 611 |
+
return {
|
| 612 |
+
"json": render_tools_json,
|
| 613 |
+
"xml": render_tools_xml,
|
| 614 |
+
"plain": render_tools_plain,
|
| 615 |
+
}[protocol](specs)
|