README.md

---
language:
- en
- rw
license: apache-2.0
task_categories:
- translation
- text-classification
tags:
- kinyarwanda
- english
- translation-quality
- quality-estimation
- human-annotations
- direct-assessment
- mt-evaluation
- african-languages
- low-resource-languages
size_categories:
- 1K<n<10K
---

# KinyCOMET Dataset — Kinyarwanda-English Translation Quality Estimation

![KinyCOMET Banner](https://huggingface.co/chrismazii/kinycomet_unbabel/resolve/main/banner.png)

## Dataset Description

This dataset contains 4,323 human-annotated translation quality assessments for Kinyarwanda ↔ English translation pairs. It was specifically created to train the [KinyCOMET model](https://huggingface.co/chrismazii/kinycomet_unbabel), addressing the critical need for reliable automatic evaluation of Kinyarwanda translations.

### Dataset Summary

- **Total Samples**: 4,323 translation pairs with quality scores
- **Languages**: Kinyarwanda (rw) ↔ English (en)
- **Annotation Method**: Direct Assessment (DA) following WMT standards
- **Annotators**: 15 linguistics students
- **Quality Control**: Minimum 3 annotations per sample, high-variance samples removed
- **Splits**: Train (3,497) / Validation (404) / Test (422)

### Why This Dataset Matters

Rwanda's MT ecosystem lacks reliable evaluation data for Kinyarwanda, a morphologically rich language where traditional metrics like BLEU correlate poorly with human judgment. This dataset provides:

- High-quality human judgments aligned with international standards
- Bidirectional coverage (both translation directions)
- Multiple MT systems evaluated (LLMs and traditional neural MT)
- Diverse domains (education, tourism)

## Dataset Structure

### Data Instances

Each sample contains:

```python
{
  'src': 'Umugabo ararya',           # Source text
  'mt': 'The man is eating',          # Machine translation
  'ref': 'The man is eating',         # Reference translation
  'score': 0.89,                      # Normalized DA score [0-1]
  'direction': 'kin2eng'              # Translation direction
}
```

### Data Fields

- **src** (string): Source text (either Kinyarwanda or English)
- **mt** (string): Machine translation output
- **ref** (string): Human reference translation
- **score** (float): Quality score normalized to [0, 1] range
  - Original scores were Direct Assessment [0-100]
  - Higher scores indicate better translation quality
- **direction** (string): Translation direction
  - `kin2eng`: Kinyarwanda → English
  - `eng2kin`: English → Kinyarwanda

### Data Splits

| Split | Samples | Percentage |
|-------|---------|------------|
| Train | 3,497 | 80% |
| Validation | 404 | 10% |
| Test | 422 | 10% |

## Dataset Creation

### Source Data

Translation pairs were sourced from three high-quality parallel corpora:

- **Mbaza Education Dataset**: Educational content
- **Mbaza Tourism Dataset**: Tourism and cultural content
- **Digital Umuganda Dataset**: General domain content

### Annotation Process

**Methodology**: Direct Assessment (DA) following WMT evaluation standards

**Annotators**: 15 linguistics students trained in translation quality assessment

**Process**:
1. Each translation pair annotated by minimum 3 different annotators
2. Annotators scored translations on 0-100 scale based on adequacy and fluency
3. Quality control: Removed 410 samples (9.48%) with standard deviation > 20
4. Final scores averaged and normalized to [0, 1] range

### Translation Systems

Six diverse MT systems were evaluated to ensure comprehensive coverage:

**LLM-based Systems**:
- Claude 3.7-Sonnet
- GPT-4o
- GPT-4.1
- Gemini Flash 2.0

**Traditional Neural MT**:
- Facebook NLLB-1.3B
- Facebook NLLB-600M

### Data Distribution

**Overall Statistics**:
- Mean score (μ): 87.73
- Standard deviation (σ): 14.14

**By Direction**:
- **English → Kinyarwanda**: μ=84.60, σ=16.28
- **Kinyarwanda → English**: μ=91.05, σ=10.47

The distribution pattern is similar to WMT datasets (2017-2022), indicating alignment with international evaluation standards.

## Usage

### Loading the Dataset

```python
from huggingface_hub import hf_hub_download
import pandas as pd

# Download dataset files
train_file = hf_hub_download(
    repo_id="chrismazii/kinycomet_dataset", 
    filename="train.csv"
)
val_file = hf_hub_download(
    repo_id="chrismazii/kinycomet_dataset", 
    filename="valid.csv"
)
test_file = hf_hub_download(
    repo_id="chrismazii/kinycomet_dataset", 
    filename="test.csv"
)

# Load the datasets
train_df = pd.read_csv(train_file)
val_df = pd.read_csv(val_file)
test_df = pd.read_csv(test_file)

print(f"Training samples: {len(train_df)}")
print(f"Validation samples: {len(val_df)}")
print(f"Test samples: {len(test_df)}")

# Convert to list of dictionaries for COMET usage
train_samples = train_df.to_dict('records')

# Example sample structure
print(train_samples[0])
```

### Using with COMET

```python
from comet import load_from_checkpoint

# Load KinyCOMET model
model = load_from_checkpoint("chrismazii/kinycomet_unbabel")

# Use your data
data = [
    {
        "src": sample['src'],
        "mt": sample['mt'],
        "ref": sample['ref']
    }
    for sample in train_samples[:10]
]

# Get predictions
segment_scores, system_score = model.predict(data, gpus=0)
print(f"System score: {system_score}")
```

## Dataset Characteristics

### Domain Coverage
- **Education**: Teaching materials, curriculum content
- **Tourism**: Travel guides, cultural information
- **General**: Mixed content from Digital Umuganda corpus

### Quality Metrics
- **Inter-annotator Agreement**: High agreement achieved through careful annotator training
- **Variance Filtering**: Samples with σ > 20 removed to ensure quality
- **Multi-annotator**: Minimum 3 annotations per sample

### Translation Direction Balance
The dataset includes both translation directions with careful attention to balance and quality:
- Adequate representation of both Kinyarwanda→English and English→Kinyarwanda
- Direction-specific evaluation possible
- Reflects real-world translation challenges in both directions

## Considerations for Using the Data

### Strengths
- High-quality human annotations following international standards
- Multiple annotators per sample for reliability
- Diverse MT systems represented
- Rigorous quality control

### Limitations
- **Domain Specificity**: Primarily education and tourism domains
- **Standard Kinyarwanda**: May not capture all dialectal variations
- **MT Systems**: Limited to six specific systems
- **Time Period**: Reflects MT quality as of 2024-2025

### Ethical Considerations
- All source data from publicly available parallel corpora
- Annotators properly compensated for their work
- No personally identifiable information included
- Open access to support African language technology development

## Additional Information

### Dataset Curators
- Jan Nehring
- Prince Chris Mazimpaka
- 15 linguistics student annotators

### Licensing
Released under Apache 2.0 License for maximum reusability.

### Citation

```bibtex
@misc{kinycomet_dataset2025,
    title={KinyCOMET Dataset: Human-Annotated Quality Estimation for Kinyarwanda-English Translation},
    author={Prince Chris Mazimpaka and Jan Nehring},
    year={2025},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/datasets/chrismazii/kinycomet_dataset}}
}
```

### Related Resources
- [KinyCOMET Model](https://huggingface.co/chrismazii/kinycomet_unbabel)
- [COMET Framework](https://unbabel.github.io/COMET/html/index.html)
- [WMT Evaluation Standards](https://www.statmt.org/wmt/)

---