Dataset mimba/text2text

📝 Description

This dataset provides multilingual parallel sentence pairs for machine translation (text-to-text tasks).
Currently, it includes Ngiemboon ↔ French (36,859 examples).
In the future, additional language pairs will be added (e.g., Ngiemboon ↔ English, etc.).

• Total examples (current): 36,859
• Columns:
  • source_text: source sentence
  • target_text: target sentence
  • source_lang: ISO 639‑3 language code of the source (e.g., nnh)
  • target_lang: ISO 639‑3 language code of the target (e.g., fra)

---

📥 Loading the dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("mimba/text2text")

print(dataset)

DatasetDict({
    nnh_fra: Dataset({
        features: ['source_text', 'target_text', 'source_lang', 'target_lang'],
        num_rows: 36859
    })
})

🔀 Train/Validation Split

The dataset is provided as a single split (nnh_fra). You can split it into train and validation/test using train_test_split:

from datasets import DatasetDict

# 90% train / 10% validation
split_dataset = dataset["nnh_fra"].train_test_split(test_size=0.1)

dataset_dict = DatasetDict({
    "train": split_dataset["train"],
    "validation": split_dataset["test"]
})

print(dataset_dict)

DatasetDict({
    train: Dataset({
        features: ['source_text', 'target_text', 'source_lang', 'target_lang'],
        num_rows: 33173
    })
    validation: Dataset({
        features: ['source_text', 'target_text', 'source_lang', 'target_lang'],
        num_rows: 3686
    })
})

⚙️ Example Usage with NLLB‑200

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "facebook/nllb-200-distilled-600M"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Add a custom language tag for Ngiemboon
tokenizer.add_tokens(["__ngiemboon__"])
model.resize_token_embeddings(len(tokenizer))

# Preprocessing
def preprocess_function(examples):
    inputs = [f"__ngiemboon__ {src}" for src in examples["source_text"]]
    targets = [tgt for tgt in examples["target_text"]]
    model_inputs = tokenizer(inputs, max_length=128, truncation=True)
    labels = tokenizer(targets, max_length=128, truncation=True)
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

tokenized_datasets = dataset_dict.map(preprocess_function, batched=True)

🌍 Available Languages

• Current:
  • nnh (Ngiemboon) ↔ fra (French)
• Planned:
  • nnh ↔ eng (English)

Additional languages to be added progressively

✅ Use Cases

• Fine‑tuning multilingual models (NLLB‑200, M2M100, MarianMT).
• Research on low‑resource languages.
• Educational demonstrations of machine translation.

BibTeX entry and citation info

@misc{
  title = {Ngiemboon ↔ French Parallel Corpus},
  author = {Mimba},
  year = {2026},
  url = {https://huggingface.co/datasets/mimba/text2text}
}

Contact For all questions contact @Mimba.