README.md

---
language:
  - bn
pretty_name: "BLUGE-BLI: Bangla NLI"
license: cc-by-4.0
task_categories:
  - text-classification
task_ids:
  - natural-language-inference
tags:
  - bengali
  - bangla
  - bluge
  - bli
  - nli
  - natural-language-inference
  - bengali-nlp
  - bnlp
  - low-resource
  - text-classification
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: "train.parquet"
      - split: validation
        path: "validation.parquet"
      - split: test
        path: "test.parquet"
---

# BLUGE-BLI: Bangla NLI

**BLUGE-BLI** is a meticulously cleaned and corrected Bangla Language Inference dataset, one of the 7 tasks in **BLUGE** (**B**engali **L**anguage **U**nderstandin**G** **E**valuation), a balanced benchmark for evaluating Bengali natural language understanding. See the full [BLUGE collection](https://huggingface.co/collections/nahid-hub/bluge) for all 7 tasks, and the [B-CORE](https://huggingface.co/datasets/nahid-hub/B-CORE-bengali-corpus) pretraining corpus and BnLM model suite released alongside it.

## Dataset Description

This task classifies premise-hypothesis sentence pairs into one of three relationships:

| Label | Relationship |
|---|---|
| `0` | Entailment |
| `1` | Neutral |
| `2` | Contradiction |

The dataset excludes premise-hypothesis pairs with fewer than 5 or more than 14 words, and addresses orthographic errors present in 3.47% of samples exhibiting surface-level issues. Corrections were proposed using GPT-4-turbo — used **exclusively** for orthographic correction, with no role in entailment/neutral/contradiction label assignment. Each proposed correction was independently reviewed and accepted, revised, or rejected by three native Bengali linguists with expertise in natural language inference (inter-annotator agreement κ=0.91), with disagreements resolved by majority vote. Of the proposals reviewed, 81.7% were accepted as-is, 12.6% were revised before acceptance, and 5.7% were rejected and reverted to the original text — semantic fidelity relative to the original texts was confirmed at 99.2% after human review.

To improve representation of underrepresented label categories, 312 additional premise-hypothesis pairs were manually created and validated through the same annotation protocol.

## Dataset Structure

Balanced **80 / 10 / 10** split:

| Split | Samples |
|---|---|
| Train | 102,767 |
| Validation | 12,845 |
| Test | 12,846 |
| **Total** | **128,458** |

**Fields:**
- `premise` — the Bangla premise sentence
- `hypothesis` — the Bangla hypothesis sentence
- `label` — inference relationship (`0`: entailment, `1`: neutral, `2`: contradiction)

## Usage

```python
from datasets import load_dataset

# Load all splits
ds = load_dataset("nahid-hub/BLUGE-bengali-nli")

# Access a specific split
train_ds = ds["train"]
val_ds   = ds["validation"]
test_ds  = ds["test"]

print(train_ds[0])
```

Load a single split directly:

```python
from datasets import load_dataset

test_ds = load_dataset("nahid-hub/BLUGE-bengali-nli", split="test")
```

Or read the Parquet files directly with pandas:

```python
import pandas as pd

train_df = pd.read_parquet(
    "hf://datasets/nahid-hub/BLUGE-bengali-nli/train.parquet"
)
```

## Annotation Process

- **Orthographic correction**: GPT-4-turbo proposals, human-reviewed only — never used for label assignment
- **Inter-annotator agreement**: κ=0.91
- **Class balancing**: 312 manually created and validated premise-hypothesis pairs
- **Semantic fidelity**: 99.2% preserved relative to original texts after correction

## License

Released under **CC BY 4.0**.

## Citation

If you use this dataset, please cite:

```bibtex
@ARTICLE{BnLM-BLUGE-B-CORE,
  author={Hossain, Nahid and Faisal Kabir, Md.},
  journal={IEEE Access}, 
  title={Efficient Monolingual Pretraining in Low-Resource Settings Through Morphology-Aware Tokenization, Principled Corpus Denoising, and Benchmark-Driven Evaluation}, 
  year={2026},
  volume={14},
  number={},
  pages={91979-92003},
  keywords={Modeling;Multilingual;Training;Cleaning;Vocabulary;Labeling;Tokenization;Computational linguistics;Pipelines;Conferences;B-CORE;BLUGE;BnLM;corpus;evaluation benchmark;pretrained models},
  doi={10.1109/ACCESS.2026.3701520}
}
```