---
language: ja
license: other
multilinguality: monolingual
tags:
- sentiment-analysis
- japanese
- e-commerce
- product-reviews
size_categories:
- 100K<n<1M
source_datasets:
- original
task_categories:
- text-classification
task_ids:
- sentiment-classification
---

# Dataset Card for bert-jp-sentiment-20250731-1035

## Dataset Summary

This dataset contains Japanese e-commerce product review sentiment data for fine-tuning BERT models.
The data combines historical training data with new human-verified sentiment labels from the RLHF feedback loop.

- **Created:** 2025-07-31T10:35:07.149235
- **Version:** 1.0
- **Total samples:** 113529

## Dataset Structure

### Data Instances

Each instance contains:
- `text`: Japanese review text
- `labels`: Sentiment label (1-5, where 1=Very Negative, 5=Very Positive)
- `source`: Data source (historical, user_feedback, ai_feedback, manual)
- `validator_id`: ID of human validator (for human-verified data)
- `verification_id`: ID of the verification record in the database

### Data Splits

| Split | Size |
|-------|------|
| train | 100914 |
| validation | 12615 |

### Label Distribution

#### Training Set
- 1 (Very Negative): 18530
- 2 (Negative): 17260
- 3 (Neutral): 25542
- 4 (Positive): 19104
- 5 (Very Positive): 20478

#### Validation Set
- 1 (Very Negative): 2316
- 2 (Negative): 2158
- 3 (Neutral): 3193
- 4 (Positive): 2388
- 5 (Very Positive): 2560

## Human Verification

- **Total human-verified samples:** 8755
- **Number of validators:** 4
- **Verification sources:** ai_feedback

## Source Distribution

- historical: 104774
- ai_feedback: 8755

## Preprocessing

The following preprocessing steps were applied:
- Text stripping and empty removal
- Encoding issue cleaning
- Stratified splitting

## Recommended Usage

### Model Training

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

# Load dataset
dataset = load_dataset("HALDATA/bert-jp-sentiment-20250731-1035" if self.organization else metadata['dataset_name'])

# The labels are 1-5. They need to be mapped to 0-4 for training.
dataset = dataset.map(lambda example: {"labels": example["labels"] - 1})

# Load recommended model
model_name = "tohoku-nlp/bert-large-japanese-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=5)

# Tokenize data
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding=True, max_length=128)

tokenized_datasets = dataset.map(tokenize_function, batched=True)
```

### Training Configuration

Recommended hyperparameters:
- Learning rate: 2e-5
- Batch size: 16
- Epochs: 3
- Max sequence length: 128
- Warmup ratio: 0.1

## Dataset Creation

### Source Data

1. **Historical Data**: Previous training datasets from production models
2. **Human-Verified Data**: New sentiment labels validated by human annotators through:
   - User feedback corrections
   - AI validation discrepancies
   - Manual annotation

### Quality Control

- Text encoding validation
- Duplicate detection and removal
- Class balance analysis
- Validator consistency checks
- Minimum samples per class verification

## Ethical Considerations

- All data has been anonymized and contains no personally identifiable information
- Human validators followed consistent annotation guidelines
- Data sources are clearly tracked for transparency

## Limitations

- Dataset is specific to Japanese e-commerce product reviews
- Sentiment categories reflect the original annotation scheme (1-5 scale)
- May contain domain-specific terminology and expressions

## License

This dataset is intended for research and development purposes.

## Citation

```
@dataset{japanese_sentiment_2025,
  title = {Japanese E-commerce Sentiment Dataset},
  author = {RLHF Training Pipeline},
  year = {2025},
  url = {https://huggingface.co/datasets/HALDATA/bert-jp-sentiment-20250731-1035}
}
```