---
language:
- en
license: mit
task_categories:
- text-classification
tags:
- sentiment
- alpaca
- sft
- instruction-tuning
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# summarization-sft-100k

Sentiment classification dataset in Alpaca instruction format for supervised fine-tuning (SFT).

**100,000 examples** sourced and normalised from:
- [mteb/mteb-human-tweet-sentiment-classification](https://huggingface.co/datasets/mteb/mteb-human-tweet-sentiment-classification)
- [stanfordnlp/sst2](https://huggingface.co/datasets/stanfordnlp/sst2)
- [SetFit/sst5](https://huggingface.co/datasets/SetFit/sst5) (5-class collapsed to 3)
- [jbeno/sentiment_merged](https://huggingface.co/datasets/jbeno/sentiment_merged)

## Label space
All labels normalised to 3 classes: `negative`, `neutral`, `positive`

## Format

| Column | Description |
|--------|-------------|
| `instruction` | Task prompt with label options |
| `input` | The text to classify |
| `output` | One of: negative, neutral, positive |

## Example

```json
{
  "instruction": "Classify the sentiment of the following text. Respond with one of: negative, neutral, positive.",
  "input": "This was an absolutely wonderful experience.",
  "output": "positive"
}
```