---
license: cc-by-sa-4.0
language:
- en
pretty_name: PART shared task (ArgMining 2027)
task_categories:
- text-classification
tags:
- argumentation
- reasoning
- argument-mining
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
- config_name: sample_test
  data_files:
  - split: test
    path: data/sample_test.jsonl
---
# PART: Parameterized Argumentative Reasoning Task — ArgMining 2027 Shared Task

Data for the PART shared task at ArgMining 2027. Task description,
submission instructions and timeline: <https://argpart.github.io/>.
Questions and submissions: [argpart2027@gmail.com](mailto:argpart2027@gmail.com).

Each item is a short puzzle: a list of witnesses, each making one statement.
Some statements are about the world ("the order was duplicated"), others say
that another witness is not credible ("Elijah is a lying"). The rule is:

> Witnesses should always be believed unless there is believable testimony
> that they are [attack word].

Two tasks are asked on the same puzzles:

- **binary:** can a given statement be believed? (`"yes"` / `"no"`)
- **multi-label:** which witnesses can be believed? (a list of names)

## Files and loading

| config | split | file | contents |
|---|---|---|---|
| `default` | `train` | `data/train.jsonl` | 3000 puzzles with answers |
| `sample_test` | `test` | `data/sample_test.jsonl` | 20 puzzles without answers, to try the submission format; **not scored** |
| `test` | `test` | `data/test.jsonl` | the real test puzzles, answers withheld (added at the test release) |

```python
from datasets import load_dataset

train = load_dataset("michiel/argpart", split="train")
sample_test = load_dataset("michiel/argpart", "sample_test", split="test")
test = load_dataset("michiel/argpart", "test", split="test")  # once released
```

The sample test set has the same format as the real one (ids like `sample_test-0001` instead of `test-0001`), so you can run your full pipeline on it and check your submission file before the test release.

## Fields

| field | splits | description |
|---|---|---|
| `id` | all | item id, e.g. `train-0001`, `sample_test-0001`, `test-0001` |
| `binary_prompt` | all | the puzzle with the binary question |
| `multi_label_prompt` | all | the same puzzle with the multi-label question (statements may be in a different order) |
| `binary` | train | answer to the binary question: `"yes"` or `"no"` |
| `multi_label` | train | names of the witnesses that can be believed |

## Submitting

Email your submission to
[argpart2027@gmail.com](mailto:argpart2027@gmail.com) with this exact
subject line, so it is not missed:

```
[PART2027 SUBMISSION] <team name> - run <n>
```

e.g. `[PART2027 SUBMISSION] Team Owl - run 1`. Attach one JSONL file per run,
named `<team>_<run>.jsonl` (e.g. `owl_1.jsonl`), with one line per test item:

```json
{"id": "test-0001", "binary": "yes", "multi_label": ["Anna", "Inez"]}
```

For questions, use the subject `[PART2027 QUESTION] <short topic>`. See
<https://argpart.github.io/> for the deadline and the number of runs allowed.

## Contact

- website (task description, instructions, timeline): <https://argpart.github.io/>
- questions about the data or the task, and submissions:
  [argpart2027@gmail.com](mailto:argpart2027@gmail.com)

## License

[CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/): you may share
and adapt the data for any purpose, including commercially, as long as you
give credit and distribute derived datasets under the same license.