vlhandfo/nordic-sentence-embedding-hard-negatives-cleaned

---
dataset_info:
features:
- name: anchor
dtype: large_string
- name: positive
dtype: large_string
- name: negative
dtype: large_string
- name: language
dtype: large_string
- name: task
dtype: large_string
- name: id
dtype: int64
- name: source
dtype: large_string
splits:
- name: train
num_bytes: 267851341
num_examples: 281938
- name: validation
num_bytes: 33501951
num_examples: 35242
- name: test
num_bytes: 33625857
num_examples: 35243
download_size: 196284042
dataset_size: 334979149
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
---

# Dataset Card for nordic-sentence-embedding-hard-negatives-cleaned

## Dataset Summary

This dataset is a cleaned and filtered triplet dataset for training sentence embedding models with hard negatives.

- **Repository**: `vlhandfo/nordic-sentence-embedding-hard-negatives-cleaned`
- **Languages**: Danish, Norwegian, Swedish
- **Final schema**: `anchor`, `positive`, `negative`, `language`, `task`, `id`, `source`
- **Splits**:
- `train`: 281,938
- `validation`: 35,242
- `test`: 35,243

Total rows: 352,423

## Source Data

The dataset is derived from:

- **Source dataset**: `DDSC/nordic-embedding-training-data`
- **Original train rows loaded**: 968,249

Each example in the final dataset carries a `source` field set to:

`DDSC/nordic-embedding-training-data`

## Preprocessing

### 1) Task Filtering (hard-negative triplets)

Only rows with non-empty text in all three triplet fields were kept:

- `query` (renamed to `anchor` in final dataset)
- `positive`
- `negative`

In practice, this removed rows with missing negatives and produced a hard-negative-only triplet set.

From the notebook run:

- Rows removed due to empty fields: 584,280
- `query_length == 0`: 14
- `positive_length == 0`: 202
- `negative_length == 0`: 584,269

After this filtering and length clipping (below), remaining task distribution is:

- `unit-triple`: 185,888
- `retrieval`: 166,535

### 2) Length Clipping

Word-length features were computed for each row:

- `query_length`
- `positive_length`
- `negative_length`

On the non-empty subset, percentile thresholds were estimated per field:

- 2.5th percentile:
- `query_length`: 2
- `positive_length`: 6
- `negative_length`: 6
- 97.5th percentile:
- `query_length`: 37
- `positive_length`: 233
- `negative_length`: 133

Rows were removed if **any** length field was outside this percentile band.

- Rows removed by clipping: 31,546

This reduces extreme short/long outliers while preserving the bulk of the distribution.

### 3) Split Generation

Data was split with stratification over `(language, task)`:

- 80% train
- 10% validation
- 10% test

using `random_state=42`.

## Data Fields

- `anchor` (`string`): Query/anchor text (renamed from source `query`)
- `positive` (`string`): Semantically relevant text
- `negative` (`string`): Hard negative text
- `language` (`string`): Language label (`danish`, `norwegian`, `swedish`)
- `task` (`string`): Task type (`retrieval` or `unit-triple`)
- `id` (`int`): Row identifier from filtered dataframe reset index
- `source` (`string`): Constant source dataset identifier