retrieval-stack/main.tex

\documentclass[10pt]{article}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage[english]{babel}
\usepackage[a4paper,margin=20mm]{geometry}
\usepackage{booktabs}
\usepackage{graphicx}
\usepackage{microtype}
\usepackage{xcolor}
\usepackage[hidelinks]{hyperref}
\usepackage{enumitem}
\Urlmuskip=0mu plus 1mu
\setlist{nosep}
\newcommand{\model}[1]{\texttt{#1}}
\title{\textbf{DUSUNEN-IR: An Evidence-First Open Stack\\for Turkish Neural Retrieval}}
\author{G\"oktu\u{g} D\"u\c{s}\"unen\\Independent Researcher, Istanbul, T\"urkiye}
\date{21 August 2026\\Author preprint}

\begin{document}
\maketitle

\begin{abstract}
Turkish neural retrieval resources are often released as isolated checkpoints,
making it difficult to trace how data selection, negative mining,
representation size, reranking, and deployment affect quality. We present
DUSUNEN-IR, an open stack connecting deterministic data preparation, dense
retrieval, model-mined negatives, Matryoshka embeddings, cross-encoder
reranking, contamination checks, and browser deployment. The release includes
100,000 training triplets, 50,000 mined hard-negative triplets, four dense
retriever configurations, two reranking studies, and a five-task Turkish MTEB
matrix. Multilingual E5-base obtains the best macro score, 0.619618. DUSUNEN
Rota v2 reaches 0.568602 and improves v1 on all five tasks, but by only 0.002706
macro points (0.48\% relative). A 118M compact retriever uses 56.1\% fewer
parameters and 40\% smaller vectors than Rota v1 with a 15.2\% relative macro
reduction. At 128 dimensions, Matryoshka training raises in-domain triplet
accuracy from 92.65\% to 94.75\% while reducing float32 vector storage by
83.3\% relative to 768 dimensions. On a frozen 102,400-pair candidate set, an
untouched multilingual cross-encoder raises MRR@10 from 0.573709 to 0.770884;
task-adapted Mercek adds only 0.001502. The stack is a reproducible resource and
a case study in publishing useful negative results without state-of-the-art
inflation.
\end{abstract}

\section{Introduction}
Dense retrieval maps queries and passages into a shared vector space for fast
nearest-neighbour search. Bi-encoders make this architecture practical
\cite{reimers2019}, while cross-encoders can improve a smaller candidate set at
higher cost \cite{nogueira2019}. Multilingual encoders and evaluation suites
have improved non-English coverage \cite{wang2022,mteb,mmteb}, but a checkpoint
alone rarely shows whether mining touched the benchmark, whether a training
gain transfers across domains, what is lost under compression, or whether a
fine-tuned reranker beats its untouched base.

DUSUNEN-IR treats data, evaluation, compression, reranking, and deployment as
one release problem. It pins external revisions, separates training and
benchmark sources, publishes raw results, and prevents model promotion when an
improvement lacks machine-readable evidence. It makes no claim of a universally
best Turkish encoder. Its empirical findings are:
\begin{enumerate}
  \item in-domain training gains need not transfer to held-out Turkish tasks;
  \item hard-negative continuation gives a small, consistent five-task gain;
  \item compact and nested embeddings expose measurable storage--quality
  operating points, but broad and in-domain evidence must remain distinct; and
  \item cross-encoder reranking gives the largest observed first-page gain,
  while further Turkish adaptation ranges from harmful to marginal.
\end{enumerate}

\section{Related work}
MS MARCO supplies large-scale ranking supervision \cite{msmarco}; mMARCO
extends it by machine translation \cite{mmarco}. E5 uses weakly supervised
contrastive pretraining and is a strong multilingual baseline \cite{wang2022}.
MTEB and MMTEB provide broad embedding evaluation \cite{mteb,mmteb}. Our five
tasks include THQuAD/TurHistQuad \cite{thquad}, XQuAD \cite{xquad}, MKQA
\cite{mkqa}, Belebele \cite{belebele}, and WebFAQ through a pinned MTEB release.
Hard negatives can strengthen decision boundaries while introducing false
negatives and miner bias \cite{simans}. Matryoshka Representation Learning
trains useful prefixes of one embedding \cite{mrl}. HNSW supplies approximate
nearest-neighbour search for mining \cite{hnsw}.

\section{Data and contamination controls}
\subsection{Retrieval 100K}
The first release derives from an Apache-2.0 Turkish MS MARCO triplet source.
The pipeline applies Unicode normalization, length and whitespace filters,
exact deduplication, deterministic SHA-256 priority sampling, and stable split
assignment. It produces 100,000 training and 2,000 validation triplets. The
source is machine translated and may contain translation artifacts.

\subsection{Hard-negative 50K}
Rota v1 encodes 50,000 training queries and searches 70,172 unique passages
originally labelled negative. HNSW search uses normalized embeddings and depth
32. Every query receives a mined negative with no fallback; mean selected cosine
is 0.550376. The released Parquet checksum is
\texttt{549c27403e51b200\allowbreak f8fa0a2200506913\allowbreak
4e673f0f89f34077\allowbreak 922894fdc5e2a457}.
The benchmark is never a mining corpus, and exact normalized overlap with
TurHistQuad is zero. This does not rule out paraphrases, semantic overlap, or
pretraining exposure.

\section{Models}
Rota v1 starts from \model{microsoft/harrier-oss-v1-270m} and trains for one
epoch on Retrieval 100K. Rota v2 continues v1 on mined negatives. Both return
640-dimensional normalized embeddings. Pusula starts from multilingual
MiniLM-L12-v2 and has 117.7M parameters with 384-dimensional output.

Atlas starts from multilingual E5-base and trains nested prefixes at 768, 512,
384, 256, 128, and 64 dimensions. Its dimension sweep uses the held-out
hard-negative split; a broad Atlas MTEB run was not performed, so these results
are separated from the main matrix.

Mercek starts from
\model{cross-encoder/mmarco-mMiniLMv2-L12-H384-v1}. An initial binary
classification run failed to beat the untouched base. The promoted v1 uses
50,000 two-document lists and LambdaLoss at $k=2$. Model selection uses training
validation only; held-out evaluation occurs afterward.

\section{Evaluation}
The official MTEB evaluator runs Turkish subsets of five pinned tasks. We
report official main scores and an unweighted task macro. Raw objects, suite
manifests, package versions, and aggregation code are public.

For reranking, Rota v2 retrieves a fixed top-100 list for each of 1,024 untouched
TurHistQuad queries. Every reranker scores the same 102,400 pairs, so metric
changes arise only from ordering. Recall@100 is fixed by construction.

Float32 storage is $4d$ bytes per vector before index overhead. Browser exports
are checked against PyTorch using cosine agreement and top-1 identity on fixed
probes.

\section{Results}
\subsection{Five-task retrieval}
\begin{table*}[t]
\centering
\small
\resizebox{\textwidth}{!}{%
\begin{tabular}{lrrrrrrrr}
\toprule
Model & Params & Dim & TurHist & XQuAD & WebFAQ & MKQA & Belebele & Macro \\
\midrule
multilingual E5-base & 278.0M & 768 & \textbf{.49726} & \textbf{.95335} & \textbf{.65032} & .07213 & \textbf{.92503} & \textbf{.619618} \\
Rota 270M v2 & 268.1M & 640 & .42198 & .86393 & .56886 & \textbf{.10331} & .88493 & \textbf{.568602} \\
Rota 270M v1 & 268.1M & 640 & .42196 & .85832 & .56402 & .10296 & .88222 & .565896 \\
Pusula 118M v0 & 117.7M & 384 & .25299 & .81123 & .46307 & .04855 & .82451 & .480070 \\
\bottomrule
\end{tabular}}
\caption{Official MTEB main scores. Macro weights each task equally.}
\label{tab:mteb}
\end{table*}

E5-base leads the suite. Rota v2 exceeds it only on MKQA, so there is no
state-of-the-art or universal-superiority claim. Hard-negative continuation
improves Rota v1 on all five tasks, but only by 0.002706 macro points. Source
hard-negative triplet accuracy remains 0.8935.

Pusula uses 56.1\% fewer parameters and 40\% smaller vectors than Rota v1, while
its macro average is 15.2\% lower relative to v1. Its source-domain triplet
accuracy rises from 0.8745 to 0.9200, illustrating domain sensitivity.

\subsection{Nested dimensions}
\begin{table}[h]
\centering
\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{rrrrrr}
\toprule
Dim & Base & Atlas & Gain & Bytes & Reduction \\
\midrule
768 & 95.25 & \textbf{95.50} & +0.25 & 3072 & 0\% \\
512 & 95.30 & \textbf{95.65} & +0.35 & 2048 & 33.3\% \\
384 & 95.05 & \textbf{95.35} & +0.30 & 1536 & 50.0\% \\
256 & 94.15 & \textbf{95.00} & +0.85 & 1024 & 66.7\% \\
128 & 92.65 & \textbf{94.75} & +2.10 & 512 & 83.3\% \\
64  & 88.75 & \textbf{92.50} & +3.75 & 256 & 91.7\% \\
\bottomrule
\end{tabular}}
\caption{In-domain triplet accuracy (\%) across embedding prefixes.}
\end{table}
Atlas improves the truncated base at every tested width. These 2,000-row
in-domain results support a storage--quality choice, not broad Turkish quality.

\subsection{Frozen-candidate reranking}
\begin{table}[h]
\centering
\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lrrrr}
\toprule
Stage & MRR@10 & nDCG@10 & R@10 & R@100 \\
\midrule
Retriever only & .573709 & .422559 & .475098 & .684082 \\
Untouched reranker & .770884 & .529506 & \textbf{.515137} & .684082 \\
\textbf{Mercek v1} & \textbf{.772386} & \textbf{.530312} & \textbf{.515137} & .684082 \\
\bottomrule
\end{tabular}}
\caption{Identical 102,400 query--passage pairs for both rerankers.}
\end{table}
The untouched cross-encoder supplies the large gain: +0.197175 MRR@10 over
retriever order. Turkish task adaptation adds +0.001502 MRR@10 and +0.000806
nDCG@10. The earlier binary run slightly reduced external metrics; both runs
are public. Recall@100 is unchanged because reranking cannot add passages.

\subsection{Deployment parity}
Rota v1's mixed-precision ONNX graph reaches 0.999956 mean cosine agreement and
1.0 top-1 agreement against PyTorch on fixed probes. Pusula's 118,335,516-byte
dynamic-int8 export reaches 0.990002 mean and 0.977650 minimum cosine agreement,
with 1.0 top-1 agreement. These are parity checks rather than cross-device
latency benchmarks.

\section{Discussion}
The experiments oppose a single leaderboard narrative. Rota v1 improves its
training-aligned validation objective yet loses to its untouched base on
TurHistQuad. Hard-negative continuation yields a consistent but small broad
gain. Pusula looks stronger on source triplets while losing on the five-task
matrix. Atlas preserves in-domain discrimination under truncation but lacks
broad evidence. Reranking produces the dominant first-page improvement while
additional task training produces only a marginal gain.

E5-base is the best default among the tested dense models when vector size is
acceptable. Rota and Pusula remain useful open Turkish experiments and
deployment targets, not replacements justified by the current matrix. Target-
corpus evaluation remains necessary.

\section{Limitations and ethics}
Training data are machine translated and can contain unnatural Turkish,
cultural distortion, or mislabeled relevance. Exact audits do not detect
semantic contamination. A mined negative may be relevant. Five tasks and one
reranking corpus omit many domains and dialects. Macro averaging can conceal
task scale. Scores are not calibrated probabilities, and retrieval does not
verify source factuality. High-stakes use requires application-specific
evaluation, source access controls, and human review.

No private documents were introduced. Public corpora are used under their
stated licences; model and data repositories retain their individual terms.

\section{Reproducibility and conclusion}
Repositories include configurations, seeds, revisions, raw MTEB objects,
candidate metrics, overlap audits, checksums, and evaluators. Initial runs used
one 8 GB RTX 5060 Laptop GPU; Atlas and Mercek v1 used one RTX 3090. Models,
data, benchmark, and demos are linked at
\url{https://huggingface.co/GoktugD}.

DUSUNEN-IR is a connected Turkish retrieval release whose claims are traceable
to public evidence. Its strongest tested dense model is an untouched baseline;
its fine-tunes expose smaller trade-offs and a marginal reranker gain. Reporting
these limits together with failed hypotheses is the practice advocated by the
stack.

\section*{Turkish summary}
DUSUNEN-IR; Turkce anlamsal arama icin veri hazirlama, dense retrieval,
hard-negative madenciligi, degisken vektor boyutlari, cross-encoder reranking,
kirlenme denetimi ve tarayicida calistirmayi acik bir arastirma zincirinde
birlestirir. Bes gorevli karsilastirmada en iyi genel model multilingual
E5-base'tir. Rota v2, Rota v1'e gore tum gorevlerde iyilesmis, ancak makro artis
yalnizca \%0.48 goreli duzeyde kalmistir. En buyuk siralama kazanci
cross-encoder asamasindan gelmistir.

\begin{thebibliography}{99}
\bibitem{reimers2019} N. Reimers and I. Gurevych. Sentence-BERT. EMNLP-IJCNLP, 2019. \url{https://arxiv.org/abs/1908.10084}
\bibitem{nogueira2019} R. Nogueira and K. Cho. Passage Re-ranking with BERT. 2019. \url{https://arxiv.org/abs/1901.04085}
\bibitem{wang2022} L. Wang et al. Text Embeddings by Weakly-Supervised Contrastive Pre-training. 2022. \url{https://arxiv.org/abs/2212.03533}
\bibitem{mteb} N. Muennighoff et al. MTEB: Massive Text Embedding Benchmark. EACL, 2023. \url{https://arxiv.org/abs/2210.07316}
\bibitem{mmteb} K. Enevoldsen et al. MMTEB: Massive Multilingual Text Embedding Benchmark. 2025. \url{https://arxiv.org/abs/2502.13595}
\bibitem{msmarco} T. Nguyen et al. MS MARCO. 2016. \url{https://arxiv.org/abs/1611.09268}
\bibitem{mmarco} L. Bonifacio et al. mMARCO. 2021. \url{https://arxiv.org/abs/2108.13897}
\bibitem{thquad} F. Soygazi et al. THQuAD. UBMK, 2021. \url{https://doi.org/10.1109/UBMK52708.2021.9559013}
\bibitem{xquad} M. Artetxe et al. On the Cross-lingual Transferability of Monolingual Representations. ACL, 2020. \url{https://arxiv.org/abs/1910.11856}
\bibitem{mkqa} S. Longpre et al. MKQA. TACL, 2021. \url{https://arxiv.org/abs/2007.15207}
\bibitem{belebele} L. Bandarkar et al. The Belebele Benchmark. ACL, 2024. \url{https://arxiv.org/abs/2308.16884}
\bibitem{simans} K. Zhou et al. SimANS. EMNLP, 2022. \url{https://arxiv.org/abs/2210.11773}
\bibitem{mrl} A. Kusupati et al. Matryoshka Representation Learning. NeurIPS, 2022. \url{https://arxiv.org/abs/2205.13147}
\bibitem{hnsw} Y. Malkov and D. Yashunin. Efficient and Robust ANN Search Using HNSW. TPAMI, 2020. \url{https://arxiv.org/abs/1603.09320}
\end{thebibliography}
\end{document}