Update README.md
Browse files
README.md
CHANGED
|
@@ -27,3 +27,15 @@ configs:
|
|
| 27 |
- split: val
|
| 28 |
path: data/val-*
|
| 29 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
- split: val
|
| 28 |
path: data/val-*
|
| 29 |
---
|
| 30 |
+
|
| 31 |
+
## Dataset Card for "vibhorag101/suicide_prediction_dataset_phr"
|
| 32 |
+
- The dataset is sourced from Reddit and is available on [Kaggle](https://www.kaggle.com/datasets/nikhileswarkomati/suicide-watch).
|
| 33 |
+
- The dataset contains text with binary labels for suicide or non-suicide.
|
| 34 |
+
- The dataset was cleaned minimally, as BERT depends on contextually sensitive information, which can worsely effect its performance.
|
| 35 |
+
- Removed numbers
|
| 36 |
+
- Removed URLs, Emojis, and accented characters.
|
| 37 |
+
- Remove any extra white spaces and any extra spaces after a single space.
|
| 38 |
+
- Removed any consecutive characters repeated more than 3 times.
|
| 39 |
+
- The rows with more than 512 BERT Tokens were removed, as they exceeded BERT's max token limit.
|
| 40 |
+
- The cleaned dataset can be found [here](https://huggingface.co/datasets/vibhorag101/phr_suicide_prediction_dataset_clean_light)
|
| 41 |
+
- The evaluation set had ~33k samples, while the training set had ~153k samples, i.e., a 70:15:15 (train:test:val) split.
|