--- license: apache-2.0 language: - en tags: - setfit - sentence-transformers - text-classification - spam-detection - email-classification - few-shot-learning - efficient - tinymodels pipeline_tag: text-classification library_name: setfit base_model: BAAI/bge-small-en-v1.5 datasets: - SetFit/enron_spam ---

### 🧠 Tiny model. Tiny dataset. Real classification. A compact **few-shot spam classifier** built with **SetFit + BAAI/bge-small-en-v1.5**.

---

## ⚡ `TINY → FAST → USEFUL`

This model is a small experiment in **few-shot text classification**. It learns to separate: > ✉️ **HAM** — legitimate email > 🚨 **SPAM** — unwanted / suspicious email The interesting part? **Only 16 labeled training examples.** ```text ┌──────────────────────┐ │ 16 EXAMPLES │ │ │ │ 8 HAM + 8 SPAM │ └──────────┬───────────┘ │ ▼ ┌───────────────────────┐ │ BAAI/bge-small-en-v1.5│ └───────────┬───────────┘ │ ▼ ┌────────────────┐ │ SetFit │ │ Few-shot NLP │ └───────┬────────┘ │ ▼ ┌───────────────────────┐ │ Logistic Regression │ └───────────┬───────────┘ │ ┌──────┴──────┐ ▼ ▼ HAM SPAM ``` --- # 📡 Model Status

| ⚙️ Component | 🔧 Configuration | | ----------------------- | ------------------------ | | **Task** | Spam Classification | | **Classes** | `2` | | **Backbone** | `BAAI/bge-small-en-v1.5` | | **Framework** | SetFit | | **Dataset** | `SetFit/enron_spam` | | **Training Examples** | `16` | | **Examples / Class** | `8` | | **Iterations** | `20` | | **Epochs** | `1` | | **Batch Size** | `16` | | **Classification Head** | Logistic Regression | | **Language** | English |

--- # 📊 Results

### `91.4%` Accuracy ### `91.4%` Macro F1
```text Accuracy ██████████████████░░ 91.4% Macro F1 ██████████████████░░ 91.4% ```

> These results come from a very small few-shot training setup. They should not be interpreted as a benchmark against production spam-filtering systems. --- # 🧬 The TinyModels Recipe ```text ┌───────────────────┐ │ SetFit/enron │ │ _spam │ └─────────┬─────────┘ │ 16 examples │ ▼ ┌───────────────────┐ │ BGE-small │ │ text encoder │ └─────────┬─────────┘ │ embeddings │ ▼ ┌───────────────────┐ │ SetFit │ │ contrastive loss │ └─────────┬─────────┘ │ ▼ ┌───────────────────┐ │ LogisticRegression│ └─────────┬─────────┘ │ ┌──────┴──────┐ ▼ ▼ ✉️ HAM 🚨 SPAM ``` ### Training configuration ```yaml backbone: BAAI/bge-small-en-v1.5 dataset: SetFit/enron_spam examples: total: 16 per_class: 8 setfit: iterations: 20 epochs: 1 batch_size: 16 loss: CosineSimilarityLoss distance_metric: cosine_distance classifier: type: LogisticRegression seed: 42 ``` --- # 🚀 Run It ```bash pip install -q setfit ``` ```python from setfit import SetFitModel model = SetFitModel.from_pretrained( "TinyModels/setfit-banking-spam" ) text = """ Congratulations! You have won $1,000,000. Click here immediately to claim your prize. """ prediction = model(text) print(prediction) ``` ### Example output ```text 1 ``` --- # 🧪 Quick Examples ### 🚨 Spam ```text CONGRATULATIONS!!! You have been selected to receive $1,000,000. Click the link below to claim your prize immediately. ``` ```text → SPAM ``` --- ### ✉️ Legitimate ```text Please find attached the global markets monitor for the week ending 12 January 2001. ``` ```text → HAM ``` --- # 🧠 Why SetFit? Traditional supervised classification can require a large labeled dataset. SetFit takes a different route: ```text Large Dataset ✕ │ │ ┌─────▼─────┐ │ SetFit │ └─────┬─────┘ │ ▼ Few labeled examples │ ▼ Useful classifier ``` This makes the experiment useful for exploring: * ⚡ Few-shot learning * 🧠 Sentence embeddings * 📚 Text classification * 🚨 Spam detection * 🔬 Efficient training * 🤗 SetFit --- # 🧩 Architecture

```text INPUT │ ▼ ┌─────────────────────────────┐ │ BAAI/bge-small-en-v1.5 │ │ │ │ Sentence Transformer │ └──────────────┬──────────────┘ │ ▼ Dense Embedding │ ▼ ┌─────────────────────────────┐ │ Logistic Regression │ └──────────────┬──────────────┘ │ ┌──────┴──────┐ ▼ ▼ HAM SPAM ✉️ 🚨 ```

--- # ⚠️ Limitations This is intentionally a **tiny experimental model**. Because only 16 examples were used for training: * Performance can vary on unseen data. * Domain shift can significantly affect predictions. * Unusual spam may be missed. * Enron-style email does not represent every modern spam pattern. * The reported score comes from a lightweight few-shot experiment. **Do not use this model as the sole component of a security-critical email filtering system.** --- # 📦 Model Identity

```text ╭──────────────────────────────────────╮ │ TINYMODEL │ ├──────────────────────────────────────┤ │ │ │ MODEL SetFit Banking Spam │ │ BACKBONE BGE-small │ │ TASK Binary Classification │ │ DATA Enron Spam │ │ EXAMPLES 16 │ │ RESULT 91.4% Accuracy │ │ │ │ STATUS ● EXPERIMENTAL │ │ │ ╰──────────────────────────────────────╯ ```

---

# ⚡ 16 Examples. # 🧠 One Small Model. # 🚨 One Real Task.

**TinyModels — building small models that actually do things.**