ChaoticEconomist/EnglishtoFrench-Translation-Dataset

---
language:
- en
- fr
license: cc-by-4.0
task_categories:
- translation
- text-generation
tags:
- translation
- french
- english
- sft
- lora
- instruction-tuning
- alpaca-format
- synthetic
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/train.csv
- split: validation
path: data/validation.csv
- split: test
path: data/test.csv
pretty_name: Eng2Fren
---

# English–French Translation Dataset (SFT / LoRA Ready)

A clean, structured dataset of **50,000 English–French sentence pairs** designed
for supervised fine-tuning (SFT) of large language models, LoRA adapters, and
general machine translation tasks.

---

## Overview

| Property | Value |
|------------------|---------------------------------|
| Language pair | English → French |
| Total rows | 50,000 |
| Train split | 45,000 (90%) |
| Validation split | 2,500 (5%) |
| Test split | 2,500 (5%) |
| Format | CSV (Alpaca-style prompt format) |
| License | CC BY 4.0 |

---

## Dataset Description

This dataset covers a wide range of everyday topics and linguistic structures,
from simple greetings and common phrases to more complex sentences involving
work, travel, health, and culture. Sentences are grammatically correct in both
English and French, with proper handling of:

- French gender agreement (masculine/feminine adjectives and articles)
- Elision rules (`Je` → `J'` before vowels, `de` → `d'` before vowels)
- Correct article usage (`Le`, `La`, `L'`, `Les`)
- BANGS adjective placement (e.g. *beau/belle* placed before the noun)
- Proper verb conjugation across first and third person
- Singular/plural agreement

---

## Columns

| Column | Type | Description |
|---------------|--------|-------------|
| `english` | string | Source sentence in English |
| `french` | string | Target translation in French |
| `instruction` | string | Task instruction: *"Translate the following English sentence into French."* |
| `prompt` | string | Alpaca-style prompt (`### Instruction / ### Input / ### Response:`) |
| `completion` | string | The French translation — the expected model output |
| `text` | string | Full `prompt + completion` string for SFT trainers |
| `source_lang` | string | Always `en` |
| `target_lang` | string | Always `fr` |
| `category` | string | Semantic category of the sentence (see below) |
| `difficulty` | string | `beginner`, `intermediate`, or `advanced` |
| `word_count` | int | Number of words in the English sentence |
| `char_count` | int | Number of characters in the English sentence |
| `split` | string | `train`, `validation`, or `test` |

---

## Categories

Sentences are labelled across 13 semantic categories:

| Category | Description |
|----------------------------|-------------|
| `greetings` | Hellos, goodbyes, introductions |
| `food_and_dining` | Restaurants, ordering, dietary preferences |
| `travel` | Transport, hotels, airports, navigation |
| `health` | Medical, pharmacy, symptoms |
| `work` | Office, jobs, meetings, contracts |
| `finance` | Banking, payments, currency |
| `technology` | Phones, internet, computers, AI |
| `education` | Study, school, books, language learning |
| `arts_and_culture` | Music, painting, photography, cinema |
| `family_and_relationships` | Family members, friendships, emotions |
| `celebrations` | Birthdays, holidays, congratulations |
| `nature_and_weather` | Seasons, environment, outdoors |
| `general` | Everyday expressions and common phrases |

---

## Difficulty Levels

| Level | Criteria |
|----------------|----------|
| `beginner` | Short sentences (≤5 words), simple vocabulary |
| `intermediate` | Medium sentences (6–10 words), everyday vocabulary |
| `advanced` | Longer sentences with complex structure |

---

## Usage

### Load with 🤗 Datasets

```python
from datasets import load_dataset

ds = load_dataset("csv", data_files="french_dataset_ml.csv")
```

### Filter by split

```python
train = ds.filter(lambda x: x["split"] == "train")
val = ds.filter(lambda x: x["split"] == "validation")
test = ds.filter(lambda x: x["split"] == "test")
```

### Fine-tune with TRL SFTTrainer

```python
from trl import SFTTrainer

trainer = SFTTrainer(
model=model,
train_dataset=train,
dataset_text_field="text", # uses the full prompt+completion field
...
)
```

### Fine-tune with Axolotl

```yaml
datasets:
- path: french_dataset_ml.csv
type: alpaca
data_files: french_dataset_ml.csv
```

### Fine-tune with LLaMA-Factory

```json
{
"dataset": "french_dataset_ml",
"dataset_format": "alpaca",
"text_field": "text"
}
```

---

## Prompt Format

All rows follow the **Alpaca instruction format**:

```
### Instruction:
Translate the following English sentence into French.

### Input:
The city is beautiful.

### Response:
La ville est belle.
```

This format is compatible with LLaMA, Mistral, Phi, Qwen, and most
instruction-tuned base models.

---

## Data Construction

Sentences were constructed and curated to ensure grammatical correctness in
both languages. The dataset prioritises **linguistic accuracy** — every sentence
has been validated for correct French grammar rules including gender agreement,
elision, verb conjugation, and adjective placement.

---

## Intended Uses

- **Translation fine-tuning** — teach a model to translate English → French
- **LoRA adapter training** — lightweight language-specific adapters
- **Instruction following** — the Alpaca format makes it suitable for general SFT
- **Benchmarking** — use the pre-split `test` set for evaluation
- **Educational tools** — language learning applications

## Out-of-Scope Uses

This dataset covers common everyday language. It is not suitable for domain-specific translation (legal, medical, technical) without augmentation.

---

## License

[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)