---
license: cc-by-4.0
language:
- de
task_categories:
- text-classification
tags:
- legal
- statutory-interpretation
- german-law
- argument-mining
pretty_name: Interpretive Canons (instance-level)
configs:
- config_name: argument
  data_files:
  - split: train
    path: arguments/argument/train.json
  - split: validation
    path: arguments/argument/validation.json
  - split: test
    path: arguments/argument/test.json
- config_name: wortlaut
  data_files:
  - split: train
    path: arguments/wortlaut/train.json
  - split: validation
    path: arguments/wortlaut/validation.json
  - split: test
    path: arguments/wortlaut/test.json
- config_name: systematik
  data_files:
  - split: train
    path: arguments/systematik/train.json
  - split: validation
    path: arguments/systematik/validation.json
  - split: test
    path: arguments/systematik/test.json
- config_name: geschichte
  data_files:
  - split: train
    path: arguments/geschichte/train.json
  - split: validation
    path: arguments/geschichte/validation.json
  - split: test
    path: arguments/geschichte/test.json
- config_name: zweck
  data_files:
  - split: train
    path: arguments/zweck/train.json
  - split: validation
    path: arguments/zweck/validation.json
  - split: test
    path: arguments/zweck/test.json
- config_name: nicht_abstrakt
  data_files:
  - split: train
    path: deutungen/nicht_abstrakt/train.json
  - split: validation
    path: deutungen/nicht_abstrakt/validation.json
  - split: test
    path: deutungen/nicht_abstrakt/test.json
- config_name: nicht_selbst_aufgestellt
  data_files:
  - split: train
    path: deutungen/nicht_selbst_aufgestellt/train.json
  - split: validation
    path: deutungen/nicht_selbst_aufgestellt/validation.json
  - split: test
    path: deutungen/nicht_selbst_aufgestellt/test.json
- config_name: konkretes_gesetz
  data_files:
  - split: train
    path: deutungen/konkretes_gesetz/train.json
  - split: validation
    path: deutungen/konkretes_gesetz/validation.json
  - split: test
    path: deutungen/konkretes_gesetz/test.json
---

# Interpretive Canons — instance-level

Companion benchmark to the paper *Classifying Interpretive Canons at the
Sentence Level: A Benchmark from the German Federal Constitutional Court*. This
is the **instance-level** release: one record per (subtask, candidate) pair,
grouped into eight subtasks, each with decision-disjoint train/validation/test
splits (no decision appears in more than one split). The companion
**decision-level** release (one record per exhaustively annotated decision, for
end-to-end pipeline evaluation) is at `felix453/interpretive-canons-decisions`.

Gold labels come from a single expert annotator. Text is German
(Bundesverfassungsgericht decisions).

## Subtasks and split sizes

The `arguments` family classifies (reading, candidate-sentence) pairs; the
`deutungen` family classifies single reading sentences.

| Subtask | Train | Validation | Test |
|---|---|---|---|
| `argument` | 40 | 40 | 120 |
| `wortlaut` | 40 | 40 | 120 |
| `systematik` | 40 | 40 | 120 |
| `geschichte` | 40 | 40 | 120 |
| `zweck` | 40 | 40 | 120 |
| `nicht_abstrakt` | 80 | 80 | 240 |
| `nicht_selbst_aufgestellt` | 80 | 80 | 240 |
| `konkretes_gesetz` | 80 | 80 | 240 |

## Record format

Each record is a JSON object `{"id", "row", "observed"}`. `row` holds the
sentence(s) and the surrounding context, with the classified sentence marked by
inline XML tags (`<deutung>...</deutung>` for readings,
`<potential_argument>...</potential_argument>` for argument candidates) inside
`text_with_context`. `observed` holds the gold labels for the subtask:

- `observed.argument` (0/1): does the candidate sentence justify the reading at all?
- `observed.wortlaut` (0/1): grammatical-canon argument?
- `observed.systematik` (0/1): systematic-canon argument?
- `observed.geschichte` (0/1): historical-canon argument?
- `observed.zweck` (0/1): objective-teleological-canon argument?
- `observed.nicht_abstrakt` (0/1): the statement is *not* abstract (concrete/case-bound).
- `observed.nicht_selbst_aufgestellt` (0/1): the statement is *not* self-advanced (cites only precedent).
- `observed.konkretes_gesetz` (list of strings): the concrete statutory provision(s) the reading interprets; empty list means none.

`observed.reasoning` is the annotator's free-text rationale where recorded.

## Loading

```python
from datasets import load_dataset
ds = load_dataset("felix453/interpretive-canons-instances", "wortlaut")
print(ds["test"][0])
```

## Raw annotations (provenance)

The raw, pre-merge Label Studio export these data are built from is available at
[`felix453/interpretive-canons-raw-export`](https://huggingface.co/datasets/felix453/interpretive-canons-raw-export).
It retains two fields the postprocessing merge does not preserve: the per-reading
determinate-content judgment `bestimmt_moeglicheDeutung` (dropped by the merge)
and the un-flattened list of concrete provisions (the merge joins the entries
into one unsplittable string). Join by Label Studio region `id`.