---
dataset_info:
  features:
  - name: sentence
    dtype: string
  - name: label
    dtype:
      class_label:
        names:
          '0': negative
          '1': positive
  - name: idx
    dtype: int32
  - name: input_ids
    sequence: int32
  - name: attention_mask
    sequence: int8
  - name: labels
    dtype: int64
  - name: task
    dtype: string
  splits:
  - name: train
    num_bytes: 49401339
    num_examples: 67349
  - name: validation
    num_bytes: 685260
    num_examples: 872
  - name: test
    num_bytes: 1425784
    num_examples: 1821
  download_size: 6173225
  dataset_size: 51512383
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# SST-2 (Modified for LLaMA 1B)
This dataset is a modified version of the **Stanford Sentiment Treebank 2 (SST-2)**, a binary classification dataset for sentiment analysis.

## **Modifications:**
- Labels were originally `"negative"` and `"positive"`, now converted to integers (`0` for negative, `1` for positive).
- Each example includes a task prefix: `Task: SST2 | Sentence: ...`
- The dataset has been **tokenized using the LLaMA-1B tokenizer**.
- Maximum sequence length is **128 tokens**.

## **Dataset Usage:**
```python
from datasets import load_dataset
dataset = load_dataset("emirhanboge/sst2_llama1b_modified")