Kinoux/french-spam-ham-detection-free-2k

French Spam/Ham Detection Free 2K

French Spam/Ham Detection Free 2K is a free 2,000-record sample extracted from the full commercial dataset French Spam/Ham Detection (56,400 synthetic messages) provided by Kinoux.

Each entry is a synthetic French message labeled with a binary classification:

• spam
• ham

The data is 100% synthetic (no personal data, no scraped emails, no real platform exports) and was generated specifically for training and evaluating French-native spam detection models.

Official product page for the full dataset and model: https://kinoux.com/datasets-models/french-spam-ham-detection/

---

Dataset Description

Summary

This dataset contains 2,000 French messages, each annotated with a single label: spam or ham.

It is designed as a free preview of the full commercial dataset (56,400 records) used to fine-tune high-performance French spam detection classifiers.

Supported Tasks

• Task: Spam detection / binary text classification
• Input: French message text (text)
• Output: Label (label) in {spam, ham}

This sample can be used to:

• Train and evaluate French spam classifiers
• Benchmark CamemBERT / Transformers pipelines on binary classification
• Build email filtering prototypes
• Create SMS or contact-form filtering systems
• Test moderation pipelines

---

Languages

• Input language: French (fr)
• Labels: English strings (spam, ham)

---

Dataset Structure

Data Instances

Each line is a JSON object like:

{"text": "Félicitations ! Vous avez gagné un iPhone 15. Cliquez ici pour réclamer votre prix.", "label": "spam"}

More examples:

{"text": "Bonjour, je vous envoie le devis demandé en pièce jointe.", "label": "ham"}
{"text": "Offre exceptionnelle valable aujourd'hui seulement ! Répondez STOP pour vous désinscrire.", "label": "spam"}
{"text": "Peux-tu me rappeler quand tu as un moment ?", "label": "ham"}

Data Fields

• text (string): Synthetic French message (email, SMS-style message, marketing text, notification, etc.).

• label (string): Binary label with two possible values:

  • spam
  • ham

---

Data Splits

This free 2K sample is provided as a single split.

Users are encouraged to create their own train/validation splits (e.g. 85/15 stratified by label).

The full 56,400-record dataset (commercial) includes a recommended stratified split and additional evaluation utilities.

---

Data Generation and Coverage

• All records are synthetically generated using French-native linguistic patterns.

• No real emails, phone numbers, or personal data are included.

• Messages simulate multiple real-world scenarios:

  • Promotional spam
  • Financial scams
  • Phishing-style messages
  • Subscription traps
  • Marketing emails
  • Legitimate personal messages
  • Business communication

The objective is to provide a clean, balanced, and legally safe corpus for training and benchmarking spam classifiers in French.

---

Usage

Example: Loading with datasets

from datasets import load_dataset

ds = load_dataset(
    "kinoux/french-spam-ham-detection-free-2k",
    split="train"
)

print(ds[0])
# {'text': '...', 'label': 'spam'}

Example: Training a CamemBERT binary classifier (sketch)

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer

dataset = load_dataset("kinoux/french-spam-ham-detection-free-2k", split="train")
dataset = dataset.train_test_split(test_size=0.15, stratify_by_column="label")

tokenizer = AutoTokenizer.from_pretrained("camembert-base")

def tokenize_batch(batch):
    return tokenizer(batch["text"], truncation=True, padding="max_length")

tokenized = dataset.map(tokenize_batch, batched=True)

label2id = {"ham": 0, "spam": 1}
id2label = {v: k for k, v in label2id.items()}

def encode_labels(batch):
    batch["labels"] = [label2id[s] for s in batch["label"]]
    return batch

tokenized = tokenized.map(encode_labels, batched=True)

model = AutoModelForSequenceClassification.from_pretrained(
    "camembert-base",
    num_labels=2,
    id2label=id2label,
    label2id=label2id,
)

args = TrainingArguments(
    output_dir="./spam_fr_model",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=3,
    weight_decay=0.01,
)

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized["train"],
    eval_dataset=tokenized["test"],
)

trainer.train()

---

Limitations

• This is a small sample (2k): suitable for experimentation and benchmarking, but not sufficient alone for production-grade filtering systems.
• Messages are synthetic; real-world performance must always be validated on in-domain data.
• Advanced evasion techniques (obfuscation, unicode tricks, heavy slang) may not be fully represented in this sample.

For production use cases, the full 56,400-record dataset is recommended.

---

Full Dataset & Model (Commercial)

The full French Spam/Ham Detection product from Kinoux includes:

• 56,400 synthetic French messages (JSONL)
• Balanced binary labels (spam, ham)
• Recommended stratified split
• Fine-tuned transformer model
• ONNX export guidance
• Clear commercial license for on-prem / private cloud usage

More details and licensing options on our website.

---

Licensing Information

This dataset is distributed under the Kinoux Dataset License (Version 2025-06-01).

• You may use this sample for internal research, prototyping, model training and evaluation.
• Redistribution of the dataset, in whole or in part, is prohibited, including uploading copies to other platforms or public repos.
• For detailed terms, please refer to the LICENSE.md file in this repository.

To obtain commercial access to the full dataset and model, please visit our website.

---

Citation

If you use this dataset in a publication, product, or internal report, please cite:

Kinoux – French Spam/Ham Detection Free 2K (synthetic dataset). Available as a sample on Hugging Face. Complete product and documentation on our website.