| ---
|
| license: apache-2.0
|
| language:
|
| - uk
|
| - en
|
| size_categories:
|
| - 10K<n<100K
|
| ---
|
| # Pretrain Dataset for Ukrainian Reranker/Embedder (80k EN→UK) |
|
|
| English retrieval dataset with pre-mined hard negatives, designed for translation to Ukrainian and use as a pretraining stage before fine-tuning on competition/domain-specific data. |
|
|
| ## Purpose |
|
|
| Stage 1 (pretrain) of a 3-stage training pipeline: |
| 1. **Pretrain** on this dataset (translated to Ukrainian) — teaches general retrieval mechanism |
| 2. **Finetune** on UA-SQuAD — adapts to Ukrainian QA patterns |
| 3. **Finetune** on competition train set — adapts to target domain (UNLP 2026) |
|
|
| ## Dataset Summary |
|
|
| | Source | Rows | Avg Positive Length | Negatives/Row | Domain | |
| |---|---|---|---|---| |
| | [Natural Questions](https://huggingface.co/datasets/tomaarsen/natural-questions-hard-negatives) | 30,000 | 611 chars | 3 | Wikipedia (broad coverage) | |
| | [HotpotQA](https://huggingface.co/datasets/sentence-transformers/hotpotqa) | 20,000 | 438 chars | 3 | Wikipedia (multi-hop reasoning) | |
| | [MS MARCO](https://huggingface.co/datasets/sentence-transformers/msmarco-co-condenser-margin-mse-sym-mnrl-mean-v1) | 20,000 | 348 chars | 1–3 | Web search (diverse queries) | |
| | [GooAQ](https://huggingface.co/datasets/tomaarsen/gooaq-hard-negatives) | 10,000 | 252 chars | 3 | Google answer boxes (lay explanations) | |
| | **Total** | **80,000** | **median 372 chars** | | | |
|
|
| ## Design Rationale |
|
|
| - **Passage length diversity**: 50-token Google snippets → 300-token multi-paragraph → 2000+ token Wikipedia sections. Matches variable chunking strategies (512-token chunks to full-page chunks). |
| - **Hard negatives only**: All sources have pre-mined hard negatives from dense retrievers or BM25 — no random/easy negatives. |
| - **Domain coverage**: Broad encyclopedic (NQ, HotpotQA), web search (MS MARCO), and informal Q&A (GooAQ) to generalize across unseen domains. |
|
|
| ## Format |
|
|
| Parquet file with columns: |
|
|
| | Column | Type | Description | |
| |---|---|---| |
| | `query` | str | Question or search query (English) | |
| | `positive` | str | Relevant passage (English) | |
| | `negative_1` | str | Hard negative passage #1 | |
| | `negative_2` | str | Hard negative passage #2 (may be empty for some MS MARCO rows) | |
| | `negative_3` | str | Hard negative passage #3 (may be empty for some MS MARCO rows) | |
| | `source` | str | Dataset origin: `nq`, `hotpotqa`, `msmarco`, `gooaq` | |
|
|
| ## Statistics |
|
|
| ``` |
| Total rows: 80,000 |
| |
| Per source: |
| gooaq : 10000 rows | avg positive length: 252 chars | has neg_1: 100% |
| hotpotqa : 20000 rows | avg positive length: 438 chars | has neg_1: 100% |
| msmarco : 20000 rows | avg positive length: 348 chars | has neg_1: 100% |
| nq : 30000 rows | avg positive length: 611 chars | has neg_1: 100% |
| |
| Query length: min=5, median=45, max=630 |
| Positive length: min=10, median=372, max=9437 |
| ``` |
|
|
| ## Translation |
|
|
| This dataset is in English. For Ukrainian pretraining, translate using: |
| - [TranslateGemma-27B](https://huggingface.co/google/translategemma-27b-it) (self-hosted, best quality for EN→UK) |
| - [Lapa LLM v0.1.2](https://huggingface.co/lapa-llm/lapa-v0.1.2-instruct) (Ukrainian-optimized, efficient tokenizer) |
| - Claude Haiku 4.5 / GPT-5.4-mini Batch API (commercial alternative) |
|
|
| Translated columns (`query_uk`, `positive_uk`, `negative_1_uk`, etc.) will be added after translation. |
|
|
| ## Intended Use |
|
|
| Training bi-encoder (embedding) and cross-encoder (reranker) models for Ukrainian document retrieval, specifically for the [UNLP 2026 Shared Task on Multi-Domain Document Understanding](https://www.kaggle.com/competitions/unlp-2026-shared-task-on-multi-domain-document-understanding). |
|
|
| Compatible with: |
| - [FlagEmbedding](https://github.com/FlagOpen/FlagEmbedding) training format (convert to `{query, pos, neg}` JSONL) |
| - [Sentence Transformers](https://sbert.net/) triplet training |
| - Any contrastive learning framework |
|
|
| ## License |
|
|
| This dataset aggregates data from multiple sources. Please refer to the original dataset licenses: |
| - Natural Questions: [Apache 2.0](https://ai.google.com/research/NaturalQuestions) |
| - HotpotQA: [CC BY-SA 4.0](https://hotpotqa.github.io/) |
| - MS MARCO: [MIT](https://microsoft.github.io/msmarco/) |
| - GooAQ: [Apache 2.0](https://github.com/allenai/gooaq) |
