TinyModels/Setfit-banking-intent

🏦 A tiny, fast, and accurate SetFit classifier that routes banking customer intents — trained with just 4 examples per class (308 sentences total).

---

🧭 Table of Contents

📖 Description  •  🏗️ Architecture  •  🚀 Quick Start  •  🏷️ Labels  •  🎓 Training  •  ⚠️ Risks  •  📜 Citation

---

📖 Model Description

🌟 Why SetFit?

SetFit achieves strong accuracy without prompts and without large labeled datasets.
It first fine-tunes a sentence embedding model with contrastive learning, then trains a lightweight classifier on top — perfect for low-data, high-class-count problems like this one.

---

🏗️ Architecture

flowchart LR
    A["📝 Input Text<br/>(customer message)"] --> B["🔤 BAAI/bge-small-en-v1.5<br/>Sentence Transformer"]
    B --> C["🧬 Sentence Embedding<br/>(384-dim)"]
    C --> D["🎯 LogisticRegression<br/>Classification Head"]
    D --> E["🏷️ Predicted Intent<br/>(1 of 77 classes)"]

    subgraph TRAIN["🎓 Training Pipeline"]
        direction TB
        T1["Contrastive Fine-tuning<br/>(CosineSimilarityLoss)"] --> T2["Embedding Extraction"]
        T2 --> T3["LogReg Head Training"]
    end

    style A fill:#6C63FF,color:#fff,stroke:#4B4BFF,stroke-width:2px
    style B fill:#FF6F91,color:#fff,stroke:#E0567A,stroke-width:2px
    style C fill:#00C2A8,color:#fff,stroke:#009E88,stroke-width:2px
    style D fill:#F7B32B,color:#000,stroke:#D99A1F,stroke-width:2px
    style E fill:#34D399,color:#000,stroke:#25A67A,stroke-width:2px
    style TRAIN fill:#1a1a2e,color:#fff,stroke:#6C63FF,stroke-dasharray: 5 5

---

🚀 Quick Start

🐍 Python — one-liner inference

pip install setfit

from setfit import SetFitModel

# 🔽 Load from the Hub
model = SetFitModel.from_pretrained("TinyModels/setfit-banking-intent")

# 🎯 Run inference
preds = model("I have an unauthorized charge.")
print(preds)
# → ['direct_debit_payment_not_recognised']

📦 Batch inference with confidence scores

from setfit import SetFitModel
import numpy as np

model = SetFitModel.from_pretrained("TinyModels/setfit-banking-intent")

texts = [
    "What do I do? My card is broken.",
    "Where should I withdraw money from?",
    "What are the currency exchange fees?",
    "I have an unauthorized charge.",
]

# Predicted labels
labels = model(texts)
print(labels)

# Probability distribution per class
probs = model.predict_proba(texts)
top_idx = np.argmax(probs, axis=1)
top_conf = probs[np.arange(len(texts)), top_idx]

for t, l, c in zip(texts, labels, top_conf):
    print(f"[{c:0.2%}] {l:<40} ← {t}")

⚡ FastAPI microservice snippet

from fastapi import FastAPI
from pydantic import BaseModel
from setfit import SetFitModel

app = FastAPI(title="Banking Intent API")
model = SetFitModel.from_pretrained("TinyModels/setfit-banking-intent")

class Query(BaseModel):
    text: str

@app.post("/classify")
def classify(q: Query):
    label = model([q.text])[0]
    return {"intent": label, "text": q.text}

---

🏷️ Model Labels (77 Classes)

💳 Card Management — 20 classes 💸 Payments — 9 classes 🔁 Transfers — 12 classes ⬆️ Top-ups — 9 classes 🏧 Cash & ATM — 8 classes 🔐 Security & PIN — 5 classes 🪪 Identity Verification — 4 classes 👤 Account Management — 3 classes 💱 Fees & Currency — 4 classes 🧾 Refunds & Disputes — 3 classes

---

🎓 Training Details

⚙️ Hyperparameters

📚 Training Set Statistics

📉 Training Curve

🧪 Framework Versions

---

⚠️ Bias, Risks & Limitations

🔍 Click to expand

🚫 Out-of-Scope Use

• ❌ Legal, medical, or financial advice generation
• ❌ Production decision-making without human-in-the-loop
• ❌ Non-English customer messages
• ❌ Emotion / sentiment / toxicity detection

---

🧩 Model Card Recipe

🎯 base_model:    BAAI/bge-small-en-v1.5
🧠 head:          LogisticRegression
📚 technique:     SetFit (few-shot contrastive)
🏷️ classes:       77
📈 metric:        accuracy
🔤 max_tokens:    512
🌱 seed:          42

---

🗺️ Roadmap

• Train on 77 banking intents
• Publish to Hugging Face Hub
• Add calibrated confidence scores
• Multilingual variant (XLM-R backbone)
• ONNX / quantized export for edge
• Evaluation on public banking benchmark

---

📜 Citation

If you use this model, please cite the SetFit paper:

@article{tunstall2022setfit,
    doi       = {10.48550/ARXIV.2209.11055},
    url       = {https://arxiv.org/abs/2209.11055},
    author    = {Tunstall, Lewis and Reimers, Nils and Jo, Unso Eun Seo and Bates, Luke and Korat, Daniel and Wasserblat, Moshe and Pereg, Oren},
    keywords  = {Computation and Language (cs.CL), FOS: Computer and information sciences},
    title     = {Efficient Few-Shot Learning Without Prompts},
    publisher = {arXiv},
    year      = {2022},
    copyright = {Creative Commons Attribution 4.0 International}
}

---

📚 Glossary

---

🌟 Star History

---

💜 Made with love by TinyModels

_{⭐ If this model helped you, consider leaving a like on the Hub/Repo!}