kalilouisangare/corpus-parallel-en-fr-bam

Global Voices EN-FR-BAM Parallel Corpus

A multilingual parallel corpus combining English, French, and Bambara, built by translating the French side of the Helsinki-NLP/opus-100 "en-fr" subset into Bambara using Facebook's NLLB-200-600M model.

---

📊 Dataset Overview

Attribute	Value
Type	Parallel Translation Corpus
Languages	English (en), French (fr), Bambara (bam)
Source	Helsinki-NLP/opus-100 — subset en-fr
Translation Model	facebook/nllb-200-distilled-600M (CT2, FP16)
Translation Direction	FR → BAM
Translation Parameters	beam_size=4, batch_size=128, max_length=128
Total Raw Lines	1,004,000
Clean Lines	853,678
Retention Rate	85.0%

---

📁 Dataset Subsets

Subset	Description	Columns	Split
all_data_not_clean	Raw 1M lines, no filtering	en, fr, bam	train=1,004,000
opus_clean_en_fr_bam	Cleaned trilingual corpus	en, fr, bam	train=850,000 / validation=1,839 / test=1,839
en_fr	English → French pairs (cleaned)	source (en), target (fr)	train=850,000 / validation=1,839 / test=1,839
fr_bam	French → Bambara pairs (cleaned)	source (fr), target (bam)	train=850,000 / validation=1,839 / test=1,839
en_bam	English → Bambara pairs (cleaned)	source (en), target (bam)	train=850,000 / validation=1,839 / test=1,839

Storage Details

Subset	Download Size	Dataset Size
all_data_not_clean	190.4 MB	286.3 MB
opus_clean_en_fr_bam	164.4 MB	246.3 MB
en_fr	119.5 MB	169.6 MB
fr_bam	107.5 MB	166.8 MB
en_bam	101.9 MB	156.1 MB

---

🧹🪣 Cleaning Pipeline

Step 1: Remove Failed Translations

• Filter: bam column is NaN (translation failed or was filtered by the pipeline)
• Removed: 44,584 lines (4.44%)
• Reason: NLLB-200 fallback mechanism or length-based filtering during translation

Step 2: Remove Exact Duplicates

• Filter: drop_duplicates(subset=['en', 'fr', 'bam'])
• Removed: 32,449 lines (3.23% of post-filtered corpus)
• Reason: Identical triplets (English, French, Bambara all identical)

Step 3: Length Ratio Alignment Filter

• FR/EN ratio: 0.5 ≤ ratio ≤ 2.0
• BAM/FR ratio: 0.5 ≤ ratio ≤ 2.5
• Removed: 73,288 lines (7.30% of post-filtered corpus)
• Reason: Structural misalignment between source and target sentences

Step 4: Extreme Length Filter

• Filter: fr ≤ 200 words AND bam ≤ 200 words
• Removed: ~0 lines (already covered by ratio filter)
• Reason: Prevents out-of-memory during training and removes truncated translations

---

📈 Corpus Statistics (Clean)

Word Count Distribution

Metric	French	Bambara	English
Mean words	15.9	16.5	14.8
Median words	9.0	9.0	9.0
P95 words	49	46	44
P99 words	79	72	70
Max words	200	200	200
Vocabulary (5k sample)	18,508	8,484	16,322

Length Ratio Distribution

Ratio	FR/EN	BAM/FR
Mean	1.134	1.164
Median	1.000	1.061
P05	0.571	0.400
P95	1.846	2.200
Aligned pairs (0.5–2.0)	95.3%	89.0%

Duplicate Analysis (Clean Corpus)

Type	Count	%
Exact triplets (EN,FR,BAM)	0	0.00%
Duplicate FR-BAM pairs	31,114	3.64%
Intentionally kept	—	—

Bucket	Count	Bar	Percentile
1-4 words	242,852	████
5-15 words	425,838	█████████████	← Median (50%)
16-30 words	209,783	██████
31-60 words	108,404	███	← P95
61-100 words	14,444	█
100+ words	2,675	░	← P99

Duplicate Analysis (Clean Corpus)

Type	Count	%
Exact triplets (EN,FR,BAM)	0	0.00%
Duplicate FR-BAM pairs	31,114	3.64%
Intentionally kept	—	—

Note: Duplicate FR-BAM pairs (e.g., "Bonjour" → "I ni sogoma") are intentionally preserved as they represent high-frequency phrases essential for NMT training.

---

Performance

Metric	Value
Throughput	~25.5 phrases/s (sustained over 4h)
Total time	~4.13 hours
Filtering rate	~4.4% (length-based during translation)
VRAM usage (inference)	~2.3 GB / 8.59 GB

---

🎯 Use Cases

Use Case	Recommended Subset	Direction
Fine-tuning NLLB-200 for FR→BAM	clean_fr_bam	fr → bam
Fine-tuning NLLB-200 for EN→BAM	clean_en_bam	en → bam
Training from scratch (multilingual)	clean_en_fr + clean_fr_bam	en ↔ fr ↔ bam
Evaluating translation quality	opus_clean_en_fr_bam (test split)	fr → bam
Research on low-resource NMT	Any cleaned subset	—
Data augmentation / back-translation	all_data_not_clean	—

from datasets import load_dataset

ds_fr_bam = load_dataset(
    "kalilouisangare/corpus-parallel-en-fr-bam", 
    "fr_bam",
    split='train', 
    streaming=True
)

ds_en_bam = load_dataset(
    "kalilouisangare/corpus-parallel-en-fr-bam",
    "en_bam",           
    streaming=True,
    split='train'
)

# Exemple FR-BAM
sample_fr = next(iter(ds_fr_bam))
print(f"1. Français : {sample_fr['source']}")
print(f"2. Bambara  : {sample_fr['target']}")
print("--"*29)
# Exemple EN-BAM
sample_en = next(iter(ds_en_bam))
print(f"3. English  : {sample_en['source']}")
print(f"4. Bambara  : {sample_en['target']}")

"""
#output:
1. Français : Aidez-moi ! Je suis seule. Je vais essayer.
2. Bambara  : A' ye ne dɛmɛ, ne kelen don. Ne bɛna a ɲini.
----------------------------------------------------------
3. English  : Help me, help me, I'm all alone now.
4. Bambara  : A' ye ne dɛmɛ, ne kelen don. Ne bɛna a ɲini.
"""

---

Limitations 👇

• Synthetic translations: Bambara side is machine-generated, not human-verified. Quality may vary for complex or rare constructions.
• Domain bias: Source corpus (Global Voices) is news/editorial — may not generalize to conversational, technical, or legal domains.
• Agglutinative language: Bambara is agglutinative; word-count based metrics may undercount morphological complexity (vocabulary appears smaller than it is).
• No human BLEU references: No professional human translations available for automatic evaluation.
• Length bias: Sentences >200 words were filtered out; very long documents are not represented.
• Translationese: The Bambara side may exhibit "translationese" artifacts from NLLB-200 (e.g., overly literal translations, calques from French syntax).

---

✍🏻 Citation

@dataset{corpus-parallel-en-fr-bam,
  title = {EN-FR-BAM Parallel Corpus},
  author = {Kalilou I Sangare},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/kalilouisangare/corpus-parallel-en-fr-bam}
}

Source Citations

• Opus-100: Zhang et al. (2020). Opus-100: An English-centric multilingual corpus for machine translation. Helsinki-NLP/opus-100
• NLLB-200: NLLB Team (2022). No Language Left Behind: Scaling Human-Centered Machine Translation. facebook/nllb-200

---

📜 License

This dataset is derived from:

• Source: Helsinki-NLP/opus-100 — license inherited from source
• Translation: Generated via Facebook's NLLB-200 (CC-BY-NC 4.0 for model weights)

Please cite both the original Opus-100 dataset and this derivative work when using this corpus.