jgalego commited on
Commit
e9f924c
·
verified ·
1 Parent(s): 59dd130

Upload folder using huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +84 -8
README.md CHANGED
@@ -15,6 +15,31 @@ tags:
15
  - pt-pt
16
  - european-portuguese
17
  - translated
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18
  ---
19
 
20
  # function-calling-pt-pt
@@ -23,10 +48,10 @@ English function-calling conversations translated into European Portuguese (pt-P
23
 
24
  | Source | License | Conversations | Kept after checks |
25
  |---|---|---|---|
26
- | [xlam](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) | cc-by-4.0 | pending | pending |
27
- | [hermes](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) | apache-2.0 | pending | pending |
28
- | [toolace](https://huggingface.co/datasets/Team-ACE/ToolACE) | apache-2.0 | pending | pending |
29
- | [when2call](https://huggingface.co/datasets/nvidia/When2Call) | cc-by-4.0 | pending | pending |
30
 
31
  The questions of the [BFCL](https://gorilla.cs.berkeley.edu/leaderboard.html) benchmark, translated the same way, are in [jgalego/bfcl-pt-pt](https://huggingface.co/datasets/jgalego/bfcl-pt-pt) for evaluation.
32
 
@@ -55,8 +80,9 @@ tools = json.loads(data[0]["tools"])
55
  ## 🔧 How it is built
56
 
57
  - Each source is parsed into one message format. Rows that do not parse are left out: Hermes' document-extraction tasks, ToolACE rows with another tool format or in Chinese, and calls that are not valid literals.
58
- - The user and assistant texts of a conversation go to [amalia-llm/AMALIA-9B-0626-DPO](https://huggingface.co/amalia-llm/AMALIA-9B-0626-DPO) in a single request, so names and terms stay the same across turns. Decoding is greedy.
59
- - Call arguments that appear in the text and are quoted, or look like a name, code or identifier, are copied as they are, as are links and emails. If the call says `"New York"`, the text says New York too; elsewhere places use their Portuguese names, such as Nova Iorque.
 
60
  - Currencies and units are not converted: "$50" becomes "50 dólares", never euros, so the text still matches the call.
61
  - A conversation is dropped if any of its translated texts fails a check:
62
 
@@ -65,12 +91,19 @@ tools = json.loads(data[0]["tools"])
65
  | `segments` | The output does not have exactly one translation per text |
66
  | `empty`, `unchanged` | A translation is empty or the same as the English |
67
  | `length` | A translation is under 0.6 or over 2 times the length of its source |
68
- | `kept` | A string that must be copied is missing |
69
  | `numbers` | A number of two or more digits is missing (an hour may move to the 24-hour clock) |
70
  | `pt-br` | It has a Brazilian form, such as "estou fazendo", "tela", "celular", "usuário" or "econômico" |
71
  | `english` | Over 15% of its words are common English words |
72
  | `too-long` | The conversation does not fit the translator's context |
73
 
 
 
 
 
 
 
 
74
 
75
  <details>
76
  <summary>The translator's instructions</summary>
@@ -98,4 +131,47 @@ Regras:
98
 
99
  ## 🙏 Credits
100
 
101
- The conversations come from [xLAM](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) (Salesforce), [Hermes Function Calling](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) (Nous Research), [ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) (Team ACE) and [When2Call](https://huggingface.co/datasets/nvidia/When2Call) (NVIDIA). Each keeps its license. The translations are by [AMALIA](https://huggingface.co/amalia-llm), a model for European Portuguese.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
  - pt-pt
16
  - european-portuguese
17
  - translated
18
+ configs:
19
+ - config_name: default
20
+ data_files:
21
+ - split: train
22
+ path:
23
+ - data/xlam.jsonl
24
+ - data/hermes.jsonl
25
+ - data/toolace.jsonl
26
+ - data/when2call.jsonl
27
+ - config_name: xlam
28
+ data_files:
29
+ - split: train
30
+ path: data/xlam.jsonl
31
+ - config_name: hermes
32
+ data_files:
33
+ - split: train
34
+ path: data/hermes.jsonl
35
+ - config_name: toolace
36
+ data_files:
37
+ - split: train
38
+ path: data/toolace.jsonl
39
+ - config_name: when2call
40
+ data_files:
41
+ - split: train
42
+ path: data/when2call.jsonl
43
  ---
44
 
45
  # function-calling-pt-pt
 
48
 
49
  | Source | License | Conversations | Kept after checks |
50
  |---|---|---|---|
51
+ | [xlam](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) | cc-by-4.0 | 59,221 | 100% |
52
+ | [hermes](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) | apache-2.0 | 1,091 | 99% |
53
+ | [toolace](https://huggingface.co/datasets/Team-ACE/ToolACE) | apache-2.0 | 10,553 | 99% |
54
+ | [when2call](https://huggingface.co/datasets/nvidia/When2Call) | cc-by-4.0 | 14,928 | 100% |
55
 
56
  The questions of the [BFCL](https://gorilla.cs.berkeley.edu/leaderboard.html) benchmark, translated the same way, are in [jgalego/bfcl-pt-pt](https://huggingface.co/datasets/jgalego/bfcl-pt-pt) for evaluation.
57
 
 
80
  ## 🔧 How it is built
81
 
82
  - Each source is parsed into one message format. Rows that do not parse are left out: Hermes' document-extraction tasks, ToolACE rows with another tool format or in Chinese, and calls that are not valid literals.
83
+ - The user and assistant texts of a conversation go to [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it) in a single request, so names and terms stay the same across turns. Decoding is greedy.
84
+ - Conversations whose user turns share a run of 10 words with a question of [jgalego/bfcl-pt-pt](https://huggingface.co/datasets/jgalego/bfcl-pt-pt) are left out, so that benchmark stays a fair test of models trained on this data.
85
+ - Call arguments that appear in the text and look like a code or identifier (digits, symbols or all capitals) are copied as they are, as are links and emails. Other words are translated: a user asks for "as luzes da cozinha" and the call still says `"kitchen"`, or for jobs in Nova Iorque while the call says `"New York"`.
86
  - Currencies and units are not converted: "$50" becomes "50 dólares", never euros, so the text still matches the call.
87
  - A conversation is dropped if any of its translated texts fails a check:
88
 
 
91
  | `segments` | The output does not have exactly one translation per text |
92
  | `empty`, `unchanged` | A translation is empty or the same as the English |
93
  | `length` | A translation is under 0.6 or over 2 times the length of its source |
94
+ | `copied` | A string that must be copied is missing |
95
  | `numbers` | A number of two or more digits is missing (an hour may move to the 24-hour clock) |
96
  | `pt-br` | It has a Brazilian form, such as "estou fazendo", "tela", "celular", "usuário" or "econômico" |
97
  | `english` | Over 15% of its words are common English words |
98
  | `too-long` | The conversation does not fit the translator's context |
99
 
100
+ | Source | Did not parse | Overlaps BFCL | Translated | segments | empty | unchanged | length | copied | numbers | pt-br | english | too-long |
101
+ |---|---|---|---|---|---|---|---|---|---|---|---|---|
102
+ | xlam | 0 | 663 | 59,337 | 0 | 0 | 3 | 0 | 4 | 26 | 6 | 77 | 0 |
103
+ | hermes | 793 | 1 | 1,099 | 0 | 0 | 0 | 0 | 0 | 8 | 0 | 0 | 0 |
104
+ | toolace | 646 | 23 | 10,631 | 0 | 0 | 6 | 0 | 3 | 48 | 4 | 17 | 0 |
105
+ | when2call | 0 | 18 | 14,982 | 0 | 0 | 0 | 0 | 0 | 26 | 1 | 27 | 0 |
106
+
107
 
108
  <details>
109
  <summary>The translator's instructions</summary>
 
131
 
132
  ## 🙏 Credits
133
 
134
+ The conversations come from the datasets below and keep their licenses. Changes: rows were parsed into one message format, filtered, and their user and assistant texts machine-translated into pt-PT; tools, calls and tool results are unchanged. The translations are by [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it).
135
+
136
+ | Source | By | License |
137
+ |---|---|---|
138
+ | [xLAM function calling 60k](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) ([APIGen](https://arxiv.org/abs/2406.18518)) | Salesforce | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
139
+ | [Hermes Function Calling v1](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) | Nous Research | [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) |
140
+ | [ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) | Team ACE | [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) |
141
+ | [When2Call](https://huggingface.co/datasets/nvidia/When2Call) (train) | NVIDIA | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
142
+
143
+ The xLAM authors release it for research and ask users to evaluate accuracy, safety and fairness before deployment; the same applies here.
144
+
145
+ If you use this data, cite the sources:
146
+
147
+ ```bibtex
148
+ @article{liu2024apigen,
149
+ title={APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets},
150
+ author={Liu, Zuxin and Hoang, Thai and Zhang, Jianguo and Zhu, Ming and Lan, Tian and Kokane, Shirley and Tan, Juntao and Yao, Weiran and Liu, Zhiwei and Feng, Yihao and others},
151
+ journal={arXiv preprint arXiv:2406.18518},
152
+ year={2024}
153
+ }
154
+
155
+ @misc{Hermes-Function-Calling-Dataset-V1,
156
+ title={Hermes-Function-Calling-Dataset-V1},
157
+ author={interstellarninja and Teknium},
158
+ url={https://huggingface.co/NousResearch/hermes-function-calling-v1}
159
+ }
160
+
161
+ @misc{liu2024toolacewinningpointsllm,
162
+ title={ToolACE: Winning the Points of LLM Function Calling},
163
+ author={Weiwen Liu and Xu Huang and Xingshan Zeng and Xinlong Hao and Shuai Yu and Dexun Li and Shuai Wang and Weinan Gan and Zhengying Liu and Yuanqing Yu and Zezhong Wang and Yuxian Wang and Wu Ning and Yutai Hou and Bin Wang and Chuhan Wu and Xinzhi Wang and Yong Liu and Yasheng Wang and Duyu Tang and Dandan Tu and Lifeng Shang and Xin Jiang and Ruiming Tang and Defu Lian and Qun Liu and Enhong Chen},
164
+ year={2024},
165
+ eprint={2409.00920},
166
+ archivePrefix={arXiv},
167
+ url={https://arxiv.org/abs/2409.00920}
168
+ }
169
+
170
+ @inproceedings{ross2025when2call,
171
+ title={When2Call: When (not) to Call Tools},
172
+ author={Ross, Hayley and Mahabaleshwarkar, Ameya Sunil and Suhara, Yoshi},
173
+ booktitle={Proceedings of NAACL 2025},
174
+ year={2025},
175
+ url={https://aclanthology.org/2025.naacl-long.174/}
176
+ }
177
+ ```