yilmazzey/sdp2-nli

---
pretty_name: SDP2 NLI (Turkish)
language:
- tr
task_categories:
- text-classification
task_ids:
- natural-language-inference
license: unknown
configs:
- config_name: multinli_tr_1_1
data_files:
- split: train
path: multinli_tr_1_1/train-*
- split: validation_matched
path: multinli_tr_1_1/validation_matched-*
- split: validation_mismatched
path: multinli_tr_1_1/validation_mismatched-*
- config_name: snli_tr_1_1
data_files:
- split: train
path: snli_tr_1_1/train-*
- split: validation
path: snli_tr_1_1/validation-*
- split: test
path: snli_tr_1_1/test-*
- config_name: trglue_mnli
data_files:
- split: train
path: trglue_mnli/train-*
- split: validation_matched
path: trglue_mnli/validation_matched-*
- split: validation_mismatched
path: trglue_mnli/validation_mismatched-*
- split: test_matched
path: trglue_mnli/test_matched-*
- split: test_mismatched
path: trglue_mnli/test_mismatched-*
dataset_info:
- config_name: multinli_tr_1_1
features:
- name: annotator_labels
list: string
- name: genre
dtype: string
- name: pairID
dtype: string
- name: promptID
dtype: string
- name: translation_annotations
struct:
- name: annotator_ids
list: int32
- name: annotator_labels
list: string
- name: translation_scores
list: int32
- name: gold_label
dtype: string
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype:
class_label:
names:
'0': entailment
'1': neutral
'2': contradiction
splits:
- name: train
num_bytes: 99665130
num_examples: 392599
- name: validation_matched
num_bytes: 3037651
num_examples: 9809
- name: validation_mismatched
num_bytes: 3175125
num_examples: 9825
download_size: 54480946
dataset_size: 105877906
- config_name: snli_tr_1_1
features:
- name: annotator_labels
list: string
- name: captionID
dtype: string
- name: pairID
dtype: string
- name: translation_annotations
struct:
- name: annotator_ids
list: int32
- name: annotator_labels
list: string
- name: translation_scores
list: int32
- name: gold_label
dtype: string
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype:
class_label:
names:
'0': entailment
'1': neutral
'2': contradiction
splits:
- name: train
num_bytes: 113121593
num_examples: 548487
- name: validation
num_bytes: 2621737
num_examples: 9836
- name: test
num_bytes: 2614627
num_examples: 9824
download_size: 26542568
dataset_size: 118357957
- config_name: trglue_mnli
features:
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype:
class_label:
names:
'0': entailment
'1': neutral
'2': contradiction
splits:
- name: train
num_bytes: 30156774
num_examples: 162788
- name: validation_matched
num_bytes: 1665313
num_examples: 9050
- name: validation_mismatched
num_bytes: 1862239
num_examples: 9200
- name: test_matched
num_bytes: 1657236
num_examples: 9008
- name: test_mismatched
num_bytes: 1853961
num_examples: 9217
download_size: 23891758
dataset_size: 37195523
---

## Dataset summary

This repository bundles **three Turkish natural language inference (NLI) datasets** into a single Hugging Face dataset repo for convenience:

- **`snli_tr_1_1`**: SNLI-TR 1.1 (Turkish translation of SNLI; originally distributed as part of NLI-TR)
- **`multinli_tr_1_1`**: MultiNLI-TR 1.1 (Turkish translation of MultiNLI; originally distributed as part of NLI-TR)
- **`trglue_mnli`**: MNLI from TrGLUE (Hugging Face: `turkish-nlp-suite/TrGLUE`, config `mnli`)

All configs are normalized to the same core schema:

- **`premise`**: string
- **`hypothesis`**: string
- **`label`**: ClassLabel with names `["entailment", "neutral", "contradiction"]`

Some configs also include additional fields from their original releases (e.g. `pairID`, `genre`, `promptID`, `captionID`, `annotator_labels`).

Note: In SNLI-TR / MultiNLI-TR 1.1, a small number of examples have missing `gold_label` (stored as `"-"`). These examples are **filtered out** during publishing so that `label` is always one of `entailment/neutral/contradiction`.

## Configs, sources, and splits

### `snli_tr_1_1`

- **Source**:
- Local copy in this project under `datasets/snli_tr_1/` (see `datasets/snli_tr_1/README.txt`)
- Upstream: Hugging Face `boun-tabi/nli_tr` and GitHub `boun-tabi/NLI-TR`
- **Splits**:
- `train`: `snli_tr_1.1_train.jsonl`
- `validation`: `snli_tr_1.1_dev.jsonl`
- `test`: `snli_tr_1.1_test.jsonl`

### `multinli_tr_1_1`

- **Source**:
- Local copy in this project under `datasets/multinli_tr_1/` (see `datasets/multinli_tr_1/README.txt`)
- Upstream: Hugging Face `boun-tabi/nli_tr` and GitHub `boun-tabi/NLI-TR`
- **Splits**:
- `train`: `multinli_tr_1.1_train.jsonl`
- `validation_matched`: `multinli_tr_1.1_dev_matched.jsonl`
- `validation_mismatched`: `multinli_tr_1.1_dev_mismatched.jsonl`

### `trglue_mnli`

- **Source**: Hugging Face dataset `turkish-nlp-suite/TrGLUE` with config `mnli`
- **Splits** (as provided by TrGLUE):
- `train`
- `validation_matched`
- `validation_mismatched`
- `test_matched`
- `test_mismatched`

## References / upstream links

- NLI-TR (SNLI-TR + MultiNLI-TR): `https://huggingface.co/datasets/boun-tabi/nli_tr` and `https://github.com/boun-tabi/NLI-TR`
- TrGLUE: `https://huggingface.co/datasets/turkish-nlp-suite/TrGLUE`

## How to load

```python
from datasets import load_dataset

snli = load_dataset("yilmazzey/sdp2-nli", "snli_tr_1_1")
multinli = load_dataset("yilmazzey/sdp2-nli", "multinli_tr_1_1")
trglue = load_dataset("yilmazzey/sdp2-nli", "trglue_mnli")
```

## Publishing / updating this repo

Use the helper script in this project:

```bash
export HF_TOKEN=***your_token_here***
./.venv/bin/python scripts/push_to_hub.py --repo-id yilmazzey/sdp2-nli
```

Notes:
- The script **reads your token from environment variables** (`HF_TOKEN` or `HUGGINGFACE_HUB_TOKEN`) and does not write it anywhere.
- Re-running the script updates the configs on the Hub.