---
language:
- en
- ja
- tr
- multilingual
license: apache-2.0
task_categories:
- text-classification
tags:
- feedback-detection
- user-satisfaction
- dialogue
- customer-service
- multilingual
- amd-mi300x
size_categories:
- 10K<n<100K
---

# Feedback Detector Dataset

A large-scale multilingual dataset for 4-class user feedback classification, labeled using GPT-OSS-120B on **AMD MI300X GPU**.

## Dataset Description

This dataset contains **51,694 examples** of user feedback classified into 4 categories:

| Label | Description | Count | % |
|-------|-------------|-------|---|
| **SAT** | User is satisfied | 8,649 | 17% |
| **NEED_CLARIFICATION** | User needs more information | 16,179 | 31% |
| **WRONG_ANSWER** | System gave incorrect response | 19,919 | 39% |
| **WANT_DIFFERENT** | User wants something different | 6,947 | 13% |

### Splits

| Split | Examples |
|-------|----------|
| Train | 46,524 |
| Validation | 5,170 |

## Data Sources

The dataset combines multiple public dialogue and complaint datasets:

| Source | Type | Examples | Languages |
|--------|------|----------|-----------|
| consumer_complaints_medium | Complaints | ~6,000 | English |
| customer_complaints | Complaints | ~2,000 | English |
| multiwoz | Task-Oriented Dialogue | ~3,000 | English |
| sgd | Schema-Guided Dialogue | ~3,000 | English |
| inscit | Information-Seeking | ~2,500 | English |
| mimics | Search Clarification | ~2,000 | English |
| hazumi | Affirmative/Negative | ~1,500 | Japanese |
| turkish_complaints | Complaints | ~1,000 | Turkish |

## Labeling Process

All examples were labeled using **OpenAI GPT-OSS-120B** served via vLLM on **AMD MI300X GPU** with:
- Temperature: 0 (deterministic)
- Structured JSON output
- Retry logic with exponential backoff
- Garbage detection and circuit breaker pattern
- Parallel processing with rate limiting

### Hardware

| Component | Specification |
|-----------|---------------|
| GPU | AMD Instinct MI300X |
| VRAM | 192 GB HBM3 |
| Framework | vLLM with ROCm |
| Model | OpenAI GPT-OSS-120B |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("llm-semantic-router/feedback-detector-dataset")

# Access splits
train = dataset["train"]
val = dataset["validation"]

# Example
print(train[0])
# {'text': '...', 'label': 0, 'label_name': 'SAT', 'source': 'multiwoz'}
```

## Schema

Each example contains:
- `text` (str): The user feedback text
- `label` (int): Numeric label (0-3)
- `label_name` (str): Label name (SAT, NEED_CLARIFICATION, WRONG_ANSWER, WANT_DIFFERENT)
- `source` (str): Original dataset source

## Label Mapping

```python
label2id = {
    "SAT": 0,
    "NEED_CLARIFICATION": 1,
    "WRONG_ANSWER": 2,
    "WANT_DIFFERENT": 3
}
```

## Intended Use

- Fine-tuning feedback detection models (ModernBERT, mmBERT, etc.)
- User satisfaction classification in conversational AI
- Customer service automation
- Dialogue system evaluation

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{feedback_detector_dataset,
  title={Feedback Detector Dataset},
  author={LLM Semantic Router Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/llm-semantic-router/feedback-detector-dataset}
}
```

## License

Apache 2.0