README.md

---
license: apache-2.0
task_categories:
- text-classification
- zero-shot-classification
language:
- en
tags:
- new_zealand
- qwen
- Qwen3Guard
pretty_name: New Zealand Classification
size_categories:
- n<1K
---

# New Zealand Guard Dataset

A binary classification dataset for training guard models to identify whether user inputs are related to New Zealand or not. This dataset is designed for fine-tuning language models to act as content filters, ensuring that only New Zealand-related queries are processed by specialised New Zealand AI assistants.

## Dataset Description

The New Zealand Guard Dataset contains **5,000 examples** of questions and statements labeled as either:
- **`related`**: Inputs that are relevant to New Zealand (cities, landmarks, famous people etc.)
- **`not_related`**: Inputs that are not related to New Zealand (general knowledge, other topics, etc.)

### Dataset Structure

Each example in the dataset follows this JSON format:

```json
{"input": "What did Kupe do?", "label": "related"}
{"input": "What is the capital of France?", "label": "not_related"}
```

### Fields

- **`input`** (string): The text input/question to be classified
- **`label`** (string): The classification label, either `"related"` or `"not_related"`

## Dataset Statistics

- **Total Examples**: 5,000
- **Format**: JSONL (JSON Lines)
- **Task**: Binary Text Classification
- **Labels**: 
  - `related`: New Zealand-related content
  - `not_related`: Non-New Zealand content

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load from Hugging Face Hub
dataset = load_dataset("geoffmunn/new-zealand-guard-dataset")

# Or load from local JSONL file
dataset = load_dataset("json", data_files="new_zealand_guard_dataset.jsonl")
```

### Example Usage in Training

This dataset is designed to be used with the Hugging Face Transformers library for fine-tuning sequence classification models. Here's a basic example:

```python
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load dataset
dataset = load_dataset("json", data_files="new_zealand_guard_dataset.jsonl")["train"]

# Map labels to IDs
LABEL2ID = {"not_related": 0, "related": 1}
ID2LABEL = {0: "not_related", 1: "related"}

dataset = dataset.map(lambda x: {"labels": LABEL2ID[x["label"]]})

# Split into train/test
dataset = dataset.train_test_split(test_size=0.1)

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B", trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
    "Qwen/Qwen3-4B",
    num_labels=2,
    id2label=ID2LABEL,
    label2id=LABEL2ID,
    trust_remote_code=True
)

# Tokenize
def tokenize_function(examples):
    return tokenizer(
        examples["input"],
        truncation=True,
        padding="max_length",
        max_length=512,
    )

tokenized_dataset = dataset.map(
    tokenize_function,
    batched=True,
    remove_columns=["input", "label"]
)
```

For a complete training script, see the reference implementation in `train_new_zealand_guard.py`.

## Use Cases

### 1. Content Moderation for New Zealand Chatbots

This dataset enables training guard models that can filter user inputs before they reach a New Zealand-specific AI assistant. Only New Zealand-related queries are allowed through, ensuring the assistant stays on-topic.

### 2. API-Based Moderation

The fine-tuned model can be deployed as a moderation API endpoint:

```python
# Example API endpoint (see new_zealand_api_server.py for full implementation)
@app.route('/api/moderate', methods=['POST'])
def moderate():
    data = request.json
    message = data.get('message', '')
    
    # Classify the message
    inputs = tokenizer(message, return_tensors="pt", truncation=True, max_length=512)
    outputs = model(**inputs)
    predicted_label = ID2LABEL[outputs.logits.argmax().item()]
    
    # Return moderation result
    risk_level = "Safe" if predicted_label == "related" else "Unsafe"
    return jsonify({
        'risk_level': risk_level,
        'predicted_label': predicted_label,
        'confidence': float(torch.softmax(outputs.logits, dim=-1).max())
    })
```

### 3. Real-Time Chat Filtering

The guard model can be integrated into chat interfaces to provide real-time moderation, blocking non-New Zealand queries before they're sent to the LLM. See `new_zealand_chat.html` for a complete implementation example.

## Model Training Recommendations

Based on the reference training script, recommended hyperparameters:

- **Base Model**: Qwen/Qwen3-4B
- **Learning Rate**: 2e-4
- **Batch Size**: 2 (with gradient accumulation of 16)
- **Epochs**: 3
- **Max Length**: 512 tokens
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
  - `r=16`
  - `lora_alpha=32`
  - `lora_dropout=0.05`
  - Target modules: `["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]`

## Dataset Examples

### Related Examples

```json
{"input": "What is Kaikoura known for?", "label": "related"}
{"input": "What tourist attractions are in Northland?", "label": "related"}
{"input": "What teams has Brendon McCullum played for?", "label": "related"}
{"input": "What is the largest lake in New Zealand?", "label": "related"}
{"input": "Who is Dame Kiri Te Kanawa?", "label": "related"}
```

### Not Related Examples

```json
{"input": "What is the capital of France?", "label": "not_related"}
{"input": "What is 2 + 2?", "label": "not_related"}
{"input": "Is the sifaka endangered?", "label": "not_related"}
{"input": "When was baseball first played?", "label": "not_related"}
{"input": "How many employees does Spotify have?", "label": "not_related"}
```

## Label Mapping

The dataset uses the following label mapping for model training:

- `"not_related"` → Class ID `0`
- `"related"` → Class ID `1`

In the context of content moderation:
- **`related`** = **Safe** (New Zealand-related content, allowed)
- **`not_related`** = **Unsafe** (Non-New Zealand content, blocked)

## Citation

If you use this dataset in your research or project, please cite it appropriately:

```bibtex
@dataset{new_zealand_guard_dataset,
  title={New Zealand Guard Dataset},
  author={Geoff Munn},
  year={2025},
  url={https://huggingface.co/datasets/geoffmunn/new-zealand-guard-dataset}
}
```

## License

Apache 2.0

## Acknowledgments

This dataset was created for training guard models to ensure New Zealand AI assistants remain focused on New Zealand-related content, improving user experience and maintaining topic relevance.

## Related Resources

- **Training Script**: See `train_new_zealand_guard.py` for a complete fine-tuning implementation
- **API Server**: See `new_zealand_api_server.py` for deployment as a moderation API
- **Chat Interface**: See `new_zealand_chat.html` for integration into a web-based chat application