Datasets:
Multi-TableGPT
Multi-TableGPT is the multilingual version of Table-GPT, the collection of table-related tasks released with the paper Table-GPT: Table-tuned GPT for Diverse Table Tasks. It is released in English, Italian, French, German and Spanish and contains about 13,000 samples in total, distributed with 40% English and 15% for each of the other four languages.
The dataset is a supervised fine-tuning (SFT) resource for table tasks: each sample is a two-turn chat, in which the user message contains the task instruction together with one or more tables in Markdown format (and, when present, the few-shot examples), and the assistant message contains the expected answer.
Source dataset
| Dataset | Table-GPT (LipengCS/Table-GPT ) |
| Paper | Table-GPT: Table-tuned GPT for Diverse Table Tasks (Li et al., 2023) |
| Generation code | microsoft/Table-GPT |
| License | MIT |
Table-GPT covers 18 table-related tasks, spanning table understanding (e.g. missing-value identification, column finding, column type annotation), table question answering, table matching (entity matching, schema matching, header-value matching), data cleaning (data imputation, error detection), data transformation (row-to-row transformation, list extraction, NL-to-SQL), data augmentation (table summarization, row and column augmentation) and table manipulation (row/column swapping, filtering and sorting). Samples are built in zero-shot and few-shot settings, may involve one or two tables, and are stored in chat format with metadata describing the task, the number of few-shot examples and the table used.
How the dataset was built
The source samples were divided into five disjoint language shares, and the non-English shares were translated through an automated translation pipeline, keeping the task definition intact:
- Language distribution: 40% of the source samples are kept in English and 15% are assigned to each of the other four languages.
- Translated content: the user and assistant turns, including the table content, with the Markdown structure of the tables preserved.
- Translation models: French and Italian were translated with
RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic; Spanish and German withQwen3-VL-235B-A22B-Instruct. - English configuration: the
enconfiguration keeps 40% of the original English samples.
Because the five configurations cover complementary portions of the source corpus, the same content is not duplicated across languages.
Dataset structure
Configurations
| Config | Language | Entries | Share |
|---|---|---|---|
en |
English | 5,279 | 40% |
fr |
French | 1,980 | 15% |
it |
Italian | 1,979 | 15% |
es |
Spanish | 1,969 | 15% |
de |
German | 1,964 | 15% |
| Total | 13,171 | 100% |
Fields
| Field | Description |
|---|---|
messages |
List of chat messages: a user turn with the task instruction, the table(s) in Markdown and the few-shot examples, and an assistant turn with the expected answer |
chat_template_kwargs |
Per-sample options for the chat template: custom_instructions, enable_thinking, python_tools and xml_tools |
lang |
Language of the sample |
task |
Table-GPT task of the sample, for example EntityMatching, RowColumnSorting or NL2SQL |
num_fewshots |
Number of few-shot examples included in the prompt |
table |
Table id used in the sample |
id |
Unique identifier of the sample |
model_version |
Translation model used for the sample (non-English configurations) |
source |
Name of the Table-GPT collection the sample comes from (English configuration only) |
Usage
from datasets import load_dataset
ds = load_dataset("VillanovaAI/Multi-TableGPT", "it", split="train")
sample = ds[0]
print(sample["task"], sample["lang"], sample["num_fewshots"])
print(sample["messages"][1]["content"][:500])
print(sample["messages"][2]["content"][:200])
Limitations
- The translations are produced automatically and may introduce errors in the instructions and in the answers.
- The language distribution is deliberately unbalanced (40% English).
Licensing and attribution
The upstream collection is released under the MIT license and this multilingual version is distributed under the same license. Content and pipeline are credited to the Table-GPT authors at Microsoft Research.
Citation
If you use this dataset, please cite the Multi-SFT paper, which describes the whole SFT collection used to train the Villanova model family, the Table-GPT paper and the dataset itself.
@misc{baiamonte2026multisft,
title={Multi-SFT: Permissively-Licensed Multilingual Instruction-Following},
author={Baiamonte, Daniela and Zeinalipour, Kamyar and Fano, Elena and Simonazzi, Stefano and Zugarini, Andrea},
year={2026},
eprint={2610.22086},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.22086}
}
@misc{li2023tablegpt,
title={Table-GPT: Table-tuned GPT for Diverse Table Tasks},
author={Li, Peng and He, Yeye and Yashar, Dror and Cui, Weiwei and Ge, Song and Zhang, Haidong and Fainman, Danielle Rifinski and Zhang, Dongmei and Chaudhuri, Surajit},
year={2023},
eprint={2310.09263},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2310.09263}
}
- Downloads last month
- -