erayalp/turkish-msmarco-listwise-relevancy

Overview

This dataset is a Turkish-translated version of the MS MARCO passage ranking data, filtered to include only queries that have two or more relevant passages. Every component (queries, passages, and structural fields) is preserved exactly as in the original dataset, with no triplet construction, no sampling, and no synthetic negatives.

The dataset is designed for IR experiments such as Listwise Reranking, SFT, Reward Modeling, and Reinforcement Learning in Turkish. Having multiple relevant passages per query provides richer supervision, allowing models to learn finer relevance distinctions and enabling more informative reward signals for reinforcement learning. This structure strengthens both listwise training and RL-based reranking by capturing more nuanced patterns in Turkish information retrieval.

---

Key Properties

• High-quality Turkish translation using Gemini 2.5 Flash
• Exactly 10 passages per query, as in the original MS MARCO reranking format
• is_selected list preserved (binary array of length 10)
• query_type, query_id, and all other metadata retained
• At least 2 relevant documents per query
• No structure modification (no triplets or custom schemas)

---

Data Fields

Each row contains:

Field	Type	Description
query	string	The Turkish-translated user query
passages	list[string]	List of 10 Turkish passages
is_selected	list[int]	Relevance labels (0/1) per passage
query_type	string	Query category (e.g., DESCRIPTION, NUMERIC, ENTITY)
query_id	int	Original MS MARCO query ID

Example slice:

• query: tavan vantilatörü kanatları hangi yönde dönmelidir
• passages: 10 translated passages (each stored as string)
• is_selected: [1, 0, 0, 0, 0, 0, 0, 0, 1, 0]
• query_type: DESCRIPTION
• query_id: 1,008,530

---

Intended Use

• Training retrieval models in Turkish
• Supervised fine-tuning using multi-relevant data
• Reward modeling and reinforcement learning with richer reward signals (GRPO, DPO, etc.)
• Low-resource IR research in Turkish
• LLM-based listwise reranking

---

Dataset Statistics

Statistic	Value
Total rows (default)	~28.2k
Train subset size	~25.1k
Train subset size	~3.1k
Passages per query	10
Avg relevant passages	≥ 2
Languages	Turkish
Translation model	Gemini 2.5 Flash

---

About This Dataset

This Turkish multi-relevant subset was translated and prepared as part of the study
“Reinforcement Learning for LLM-Based Listwise Reranking in Turkish Information Retrieval”
conducted by Eray Alp & Tunahan Ibiş.

The study has not been formally published yet.

---

License

This dataset is a translated derivative of MS MARCO.
You must comply with Microsoft’s MS MARCO license.

---

Citation

Original MS MARCO Dataset

@article{DBLP:journals/corr/NguyenRSGTMD16,
  author    = {Tri Nguyen and
               Mir Rosenberg and
               Xia Song and
               Jianfeng Gao and
               Saurabh Tiwary and
               Rangan Majumder and
               Li Deng},
  title     = {{MS} {MARCO:} {A} Human Generated MAchine Reading COmprehension Dataset},
  journal   = {CoRR},
  volume    = {abs/1611.09268},
  year      = {2016},
  url       = {http://arxiv.org/abs/1611.09268},
  archivePrefix = {arXiv},
  eprint    = {1611.09268},
  timestamp = {Mon, 13 Aug 2018 16:49:03 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/NguyenRSGTMD16.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}