Makan09/Bambara_Translation_Corpus_FR-BM

🌍 Bambara Translation Corpus: Direct & Instruction-Tuned (FR-BM)

🚀 Overview & Vision

The Bambara Translation Corpus is a comprehensive bilingual dataset designed to bridge French and Bamanankan across two distinct paradigms: Direct Neural Machine Translation (NMT) and Prompt-Based Instruction Tuning.

This dual-mode architecture caters to both traditional seq2seq translation models and modern instruction-tuned Large Language Models (LLMs).

---

📊 Dataset Formats & Modes

The corpus supports two structural configurations to accommodate different model training workflows:

1. Direct Translation Mode (Without Instruction)

Ideal for standard NMT models (e.g., MarianMT, NLLB, T5) or simple sequence-to-sequence fine-tuning.

{
  "src": "Bonjour, comment allez-vous ?",
  "tgt": "I ni sɔgɔma, i ka kɛnɛ wa?"
}

2. Instruction-Guided Mode (With Instruction)

Designed for Supervised Fine-Tuning (SFT) of LLMs (e.g., LLaMA, Mistral, Gemma) to enable conversational, prompt-driven translation agents.

{
  "instruction": "Traduis le texte suivant du français vers le bambara.",
  "input": "Bonjour, comment allez-vous ?",
  "output": "I ni sɔgɔma, i ka kɛnɛ wa?"
}

---

💡 What This Dataset Can Achieve

• Dual Model Compatibility: Train lightweight, dedicated translation engines alongside prompt-following LLM assistants using the same core alignment data.
• Prompt-Based Translation Agents: Enable conversational AI to dynamically process translation directives (e.g., formal vs. informal phrasing, domain-specific terminology).
• Cultural & Contextual Alignment: Preserve nuance, idioms, and natural phrasing native to Malian speech patterns in both direct and conversational modes.

---

💻 Quick Start (Python)

You can easily load and explore the dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("Makan09/
Bambara_Translation_Corpus_FR-BM")

# Inspect a direct translation pair vs an instruction-guided pair
print("Sample:", dataset["YB-bamanankan_instruction_(in-out)v1.jsonl"][0])

---

Maintained independently by Makan09.