Turkish AI · security models · measured systems

I build Turkish AI.
End to end.

I'm Göktuğ. I turn raw data into trained models, reproducible benchmarks and working inference products. Every claim links back to an artifact you can inspect, run and measure.

Göktuğ Düşünen
100Kcurated train triplets
50Kmined hard negatives
5held-out Turkish tasks
0exact benchmark overlaps
01 / Flagship stack

DUSUNEN Model Lab.

A measured Turkish AI program: variable-dimension embeddings, compact retrieval, learning-to-rank, pinned benchmarks and working browser inference.

MODEL / MATRYOSHKA RETRIEVER

DUSUNEN Atlas 278M

One Turkish retrieval model with six useful embedding sizes. At 128 dimensions it cuts index storage by 83.3% while reaching 94.75% held-out hard-negative accuracy.

768→64-DMATRYOSHKABF16
MODEL / COMPACT RETRIEVER

DUSUNEN Pusula 118M

A 117.7M encoder with six selectable dimensions, a 91.7% index-size reduction at 32-D, and a pinned five-task Turkish scorecard.

384→32-DMATRYOSHKA5-TASK MTEB
MODEL / CROSS-ENCODER

DUSUNEN Mercek 118M

LambdaLoss learning-to-rank over 50K lists, with measured MRR@10 and nDCG@10 gains over its untouched multilingual base on 102,400 frozen pairs.

LAMBDA LOSS102.4K PAIRSHELD-OUT UPLIFT
BENCHMARK / FIVE TURKISH TASKS

One protocol. Four systems. Raw evidence.

TurHistQuad, XQuAD, WebFAQ, MKQA and Belebele through the official MTEB evaluator, with pinned revisions and explicit model prompt formats.

MTEBNDCG@10REPRODUCIBLE
DATASET / MINED NEGATIVES

50K difficult training triples

Model-mined negatives from a 70K candidate pool, with lexical safeguards, zero fallbacks, checksums and an exact-overlap audit.

HNSWPARQUET0 EXACT OVERLAP
SPACE / CITED RETRIEVAL

Ask in Turkish. Inspect the sources.

A free browser-only retrieval demo that selects extractive evidence and links every answer fragment to its source. No hosted model API and no hidden generation step.

CITED RAGTRANSFORMERS.JS0 PAID API
SPACE / VECTOR BUDGET LAB

Watch retrieval change from 384-D to 32-D

One browser query, six live Matryoshka vector budgets, instant ranking changes and explicit index-size trade-offs.

384→32-DONNX Q8LOCAL INFERENCE
02 / Security AI

NanoSOC1:8B.

A gated Foundation-Sec 8B QLoRA system for structured SOC triage, evidence correlation, MITRE ATT&CK attribution and human-approved response—published with both its benchmark gains and its failure modes.

03 / Pipeline

No missing middle.

The work connects data decisions to measured model behavior and finally to something people can use.

Curate

Normalize, filter, deduplicate, checksum and publish the exact training split.

Train

Contrastive fine-tuning with memory-aware batches on a single 8 GB GPU.

Mine

Retrieve difficult negatives, reject risky matches and record every mining decision.

Evaluate

Pinned held-out data, exact vector search, strong baselines and raw JSON results.

Ship

Open weights, documented inference and a live semantic-search experience.

04 / Engineering

Built for scrutiny, not screenshots.

ProvenanceLicenses · revisions · checksums
TrainingPyTorch · CUDA · BF16
EvaluationFAISS · MRR · nDCG · recall
DeliveryHF Hub · ONNX · Static Spaces