README.md

---
language:
- pl
- en
pretty_name: FIQA PL/EN Reranked Hard Negatives
task_categories:
- sentence-similarity
tags:
- retrieval
- text-embeddings
- hard-negatives
- reranking
- fiqa
- polish
- english
configs:
- config_name: pl
  default: true
  data_files:
  - split: train
    path: train_pl.jsonl
- config_name: en
  data_files:
  - split: train
    path: train_en.jsonl
---

# FIQA PL/EN Reranked Hard Negatives

This dataset contains two aligned JSONL training files for embedding-model fine-tuning:

- `train_pl.jsonl`: Polish texts.
- `train_en.jsonl`: English texts.

Both files contain the same 5,500 training records in the same order. The text fields (`query`, `pos`, `neg`) are language-specific, while all metadata fields are shared across the aligned Polish and English examples. This means that the same query-document pair has the same IDs, reranker scores, retrieval scores, and selection metadata in both language versions.

The Polish subset is configured as the default Dataset Viewer subset.

## Dataset Statistics

| Metric | Value |
| --- | ---: |
| Queries per language | 5,500 |
| Total positives per language | 14,850 |
| Negatives per query | 10 |
| Queries with mined synthetic positives | 684 |
| Mined synthetic positives | 684 |
| Positives with `pos_scores_stronger_reranker > 23.50` | 3,190 |
| Queries with at least one positive above `23.50` | 2,086 |
| Queries with no positive above `23.50` | 3,414 |
| Mean `neg_scores` | -3.5121 |
| Mean `neg_scores` for queries with at least one positive above `23.50` | -3.4766 |

## Text Length Statistics

Character lengths are measured in characters. Token lengths include model special tokens and were computed without padding or truncation.

### Character Lengths

| Language | Texts | Count | Mean | Median | p90 | p95 | p97 | p99 |
| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| EN | Queries | 5,500 | 61.5 | 59 | 92 | 103 | 109 | 124 |
| EN | Positive passages | 14,850 | 1,019.7 | 779 | 2,050 | 2,707 | 3,157 | 4,389 |
| EN | Negative passages | 55,000 | 640.2 | 409 | 1,339 | 2,041 | 2,598 | 3,678 |
| PL | Queries | 5,500 | 72.4 | 69 | 111 | 124 | 134 | 153 |
| PL | Positive passages | 14,850 | 1,086.5 | 840 | 2,158 | 2,846 | 3,319 | 4,508 |
| PL | Negative passages | 55,000 | 683.8 | 436.5 | 1,406 | 2,252 | 2,798 | 3,708 |

### Token Lengths

| Tokenizer | Language | Texts | Count | Mean | Median | p90 | p95 | p97 | p99 |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `sdadas/mmlw-retrieval-roberta-base` | EN | Queries | 5,500 | 23.5 | 23 | 35 | 39 | 41 | 47 |
| `sdadas/mmlw-retrieval-roberta-base` | EN | Positive passages | 14,850 | 352.5 | 270 | 704 | 929 | 1,084 | 1,515 |
| `sdadas/mmlw-retrieval-roberta-base` | EN | Negative passages | 55,000 | 226.1 | 144 | 474 | 725 | 912 | 1,277 |
| `sdadas/mmlw-retrieval-roberta-base` | PL | Queries | 5,500 | 16.2 | 15 | 24 | 27 | 30 | 34 |
| `sdadas/mmlw-retrieval-roberta-base` | PL | Positive passages | 14,850 | 221.2 | 171 | 438 | 574 | 671 | 934 |
| `sdadas/mmlw-retrieval-roberta-base` | PL | Negative passages | 55,000 | 147.2 | 94 | 302 | 469 | 575 | 852 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | EN | Queries | 5,500 | 21.1 | 20 | 31 | 34 | 36 | 41 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | EN | Positive passages | 14,850 | 310.0 | 237 | 618 | 819 | 952 | 1,329 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | EN | Negative passages | 55,000 | 199.5 | 128 | 421 | 642 | 800 | 1,142 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | PL | Queries | 5,500 | 15.5 | 15 | 23 | 26 | 28 | 31 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | PL | Positive passages | 14,850 | 207.1 | 161 | 409 | 536 | 623 | 866 |
| `sdadas/mmlw-retrieval-roberta-large-v2` | PL | Negative passages | 55,000 | 137.1 | 88 | 281 | 438 | 536 | 784 |

## Data Construction

This dataset was prepared with the help of [lightonai/embeddings-fine-tuning](https://huggingface.co/datasets/lightonai/embeddings-fine-tuning). In the LightOnAI dataset, the authors computed 2,048 document candidates for each query using [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base). The similarity scores produced by LightOnAI are stored in `original_pos_scores` and `original_neg_scores`.

For this prepared FIQA export, candidate query-document pairs were reranked with [mixedbread-ai/mxbai-rerank-base-v2](https://huggingface.co/mixedbread-ai/mxbai-rerank-base-v2). The resulting scores are stored in `pos_scores` and `neg_scores`. Reranking was performed until 10 negatives satisfying the selection rules were found for each query. During the same process, if candidate documents met the positive thresholds, they were added as mined positives.

At the end of the pipeline, positives were also scored with [BAAI/bge-reranker-v2.5-gemma2-lightweight](https://huggingface.co/BAAI/bge-reranker-v2.5-gemma2-lightweight). These scores are stored in `pos_scores_stronger_reranker`. This stronger reranker score is intended for quality filtering before training: original datasets can contain query-positive pairs where the positive passage does not actually answer the query, so positives with too-low `pos_scores_stronger_reranker` values should be removed before model training.

## Dataset Structure

Each JSONL row has the following structure:

```json
{
  "query_id": "448",
  "query": "...",
  "pos": ["..."],
  "neg": ["..."],
  "pos_scores": [5.75],
  "neg_scores": [-2.5, -2.5],
  "prompt": "",
  "type": "retrieval"
}
```

The actual records contain all fields documented below.

## Text Fields

| Field | Description |
| --- | --- |
| `query` | Query text in the language of the subset. |
| `pos` | List of positive passages/documents in the language of the subset. |
| `neg` | List of negative passages/documents in the language of the subset. |

## Metadata Fields

| Field | Description |
| --- | --- |
| `query_id` | Query/record identifier used to align the Polish and English versions. |
| `pos_scores` | Scores for `query` + positive pairs computed with `mixedbread-ai/mxbai-rerank-base-v2`. |
| `neg_scores` | Scores for `query` + negative pairs computed with `mixedbread-ai/mxbai-rerank-base-v2`. |
| `prompt` | Training prompt field; empty in these files. |
| `type` | Task type; `retrieval` in these files. |
| `pos_id` | Document IDs for positives. |
| `neg_id` | Document IDs for negatives. |
| `pos_is_synthetic` | Boolean flags indicating whether each positive was mined by the pipeline. `false` means the positive came from the original FIQA dataset; `true` means it was mined and should be treated as synthetic. |
| `neg_selection_tier` | Negative-selection label for each negative. Negatives selected from the 2,048 LightOnAI candidates keep their tier label, such as `strict` in this export. If the 2,048 candidates did not provide enough valid negatives, additional documents were sampled from the full corpus and the highest-scoring valid ones were added with `relaxed_backfill`. |
| `original_pos_scores` | LightOnAI similarity scores for positives, computed with `Alibaba-NLP/gte-modernbert-base`. |
| `original_neg_scores` | LightOnAI similarity scores for negatives, computed with `Alibaba-NLP/gte-modernbert-base`. |
| `pos_scores_stronger_reranker` | Positive scores computed at the final stage with `BAAI/bge-reranker-v2.5-gemma2-lightweight`; intended for filtering low-quality positives before training. |
| `mean_neg_score` | Mean value of `neg_scores` for the record. |
| `mean_pos_score` | Mean value of `pos_scores` for the record. |
| `max_neg_score` | Maximum value of `neg_scores` for the record. |
| `difference_between_mean_scores` | Difference between `mean_pos_score` and `mean_neg_score`. |
| `difference_between_max_scores` | Difference between `mean_pos_score` and `max_neg_score`. |

## Intended Use

The dataset is intended for training and evaluating retrieval or embedding models with query-positive-negative examples. The aligned Polish and English subsets can be used for monolingual training, bilingual comparisons, or multilingual fine-tuning setups.

Before training, consider filtering positives using `pos_scores_stronger_reranker` to remove weak query-positive pairs.