TasaractDavut Emre Taşar

Emre Tesseract

Welcome. Sometimes na the tone for one Turkish sentence catch my ear. Sometimes na wetin one quantum measurement fit tell us. See where those questions carry me go.

Scroll down to see more You dey travel through my research

One model fit change another model weights?

I test changes to the weights of one classifier wey get permanent damage, with the UCI Adult dataset. The change wey Sol suggest improve balanced accuracy, but overall accuracy drop and the probability estimates come less reliable. For the intact SmolLM2 language model, every selection for ARC-Easy and PIQA keep the original weights.

Classifier wey get damage

UCI Adult · Sol

Starting pointWetin dem select
Accuracy wey balance the classes78.59%79.02%
Overall accuracy82.86%82.63%

Language model wey no get damage

SmolLM2-1.7B · ARC-Easy / PIQA

Na the original weights dem select

Measurements · GitHub
Emre stand beside one Elektra espresso machine
Na dis espresso machine I like pass.
Emre dey explain im PhD research for Navarra
One photo from wen I present my PhD research for di fifth DATAI scientific meeting at University of Navarra.
Emre hold di BBVA We Think Big award
For 2024, BBVA give me the We Think Big award for my work on patents and wetin I contribute to academic research.
Emre dey smile for construction site. E wear white hard hat and bright safety vest.
Na from di time wey I bin dey work as civil engineer.

Hello, na me be Emre.

I come from İzmir. I bin be civil engineer. Now I dey work as data scientist for BBVA AI Innovation team for Madrid. I dey learn Spanish too as I dey get to know di city.

I dey do my PhD for University of Navarra. I dey study how banks fit make fairer decisions if dem measure and reduce di uncertainty for wetin their AI models predict. Abstract maths, quantum physics and philosophy dey interest me too. I like how one question fit carry me from one of dem enter another.

For house I dey make coffee well-well. I like to paint, play chess and make up funny stories for my daughter. I get one small darbuka, one kalimba and one Zanpakuto too.

How to reach me and see my profiles

Hugging Face / GitHub

Models and code

See wetin each public model or code repository dey do. Each one get link to the documentation.

Hugging Face

Turkish Sentiment Analysis · TSAM

I train dis BERT model to pick di tone wey Turkish review get. E dey give separate score for positive, neutral and negative tone.

12,053Total downloads for di Hub18 Nọv 2026
GitHub

Regression Conformal Prediction Research Atlas

Use the atlas compare prediction intervals on 67 datasets. You fit see how wide each interval be and how often the actual result fall inside am.

GitHub

Quantum Time Experiment

The experiments check whether records wey different parts of a quantum system keep fit show the order wey events happen, and when another reference dey necessary.

Hugging Face

Whisper Medium TR

Give the model a recording in Turkish and e go write out the speech. I adapt Whisper Medium with Turkish recordings from Mozilla Common Voice 11.0.

Hugging Face

Gemma 2 · Turkish Wikipedia

This Gemma 2 9B adapter learn further from Turkish Wikipedia and fit continue a Turkish passage. The model card allow research and academic use and limit commercial use to BBVA Group.

Hugging Face

Java Methods · Doc2Vec

This model turn a Java method into 200 numbers wey represent the code. Developers fit use dem to find code wey resemble one another or to group methods.

Hugging Face

Gemma 3 27B · Turkish reasoning

I fine-tune the Unsloth 4-bit version of Gemma 3 27B Instruct to answer reasoning questions in Turkish. The card include early TARA results for logic, planning and SQL.

Hugging Face

arXiv Abstract-to-Title · mT5

Give this mT5-small adaptation a research abstract and e suggest a title for the paper. E learn from arXiv abstract and title pairs.

Hugging Face

Wav2Vec2 XLS-R · Turkish

This adaptation of Wav2Vec2 XLS-R turn recorded Turkish speech to writing. The model card show how the training go and how often e make word errors.

Hugging Face

Qwen 0.5B · GRPO maths

Give this small Qwen model a maths word problem. Its GSM8K training score the numerical answer and the way e arrange the solution separately.

GitHub

BIRD · Text-to-SQL experiments

Give the pipeline a question and the database structure. E generate SQL, run the query and check wetin come out.

GitHub

WebCLI

Search public technical documents on your own machine from the command line. WebCLI give stable references and fit export the results as JSONL.

GitHub

Fairness Metrics

Two groups fit get positive decisions at the same rate, and the kinds of mistakes wey affect dem fit still differ. This Python example use user ID join predictions, actual results and group membership.

GitHub

PaperBanana field test

I use the methods sections from three of my papers to test PaperBanana. You fit compare each picture wey e generate with the original and see where e miss the mark.

GitHub

World Cup 2026 Nowcast Lab

This experiment use probabilities for each match to simulate the 2026 World Cup and estimate the chance wey each team get to win the tournament.

GitHub

Agent Chronicle

The tool arrange coding-agent decisions, file changes and checks as dem happen. You fit trace a progress claim back to the activity wey support am.

Map of my research

Map of my research

Choose one research area

When model no sure

How model fit know the limit of wetin e sabi?

I use conformal prediction to measure how often the right answer fall inside the set wey model give, and to decide when person need step in.

Graph wey link ideas

Four questions from my notebook

Ontologies and data structuresWetin go change for graph if one train come late?

Station na place. Train departure na event. If I use different node type for each one and keep where each time come from, I fit trace all the connections wey one delay go affect?

RDF 1.1 · W3C
Meaning-Native ComputingIf computer know the rules wey one program must follow, e fit predict better which instructions go run next?

Meaning-Native Computing na my own independent research project. I wan compare two ways to predict wetin go happen, with the same program and about the same computing power. I go give one of dem the program rules too, write am out clear. E go help am see the next steps better? Before any change wey e predict fit take effect, another check must confirm say the change follow the rules.

Category theoryWetin go remain when I join two transformations?

One process turn timetable to graph. Another one find route for the graph. When I join the two, how I fit keep track of which time belong to which station?

Seven Sketches in Compositionality
Quantum measurementWetin go happen if I ask the same qubit another question?

I prepare the “+” state. If I measure for X basis, na + I go always get. For Z basis, 0 and 1 get equal chance. Wetin dey fascinate me be say the answer dey change with how I measure am, even when I prepare am the same way.

Quantum Information · IBM
TÜRKPATENT / 2022–2026

Patents and utility models

My name dey as inventor for 30 applications wey Türkiye Garanti Bankası A.Ş. file for Türkiye. Na 28 patent applications and two utility model applications.

30Applications
Drawing of di idea · Small robot for search and rescue
Drawing of di ideaWO2025071525A1 · Fig. 1

TR 2023/011901

Small robot for search and rescue

Original documents

Türkiye don announce the final decision to grant this patent ·

TR 2023/004244

Control POS terminal with AI

Original documents
Check the register (30)

Official register · For the current legal status of any application, check the original register.

Research note / 2025

Whose case dem go check again?

I use six datasets compare two ways to set the model threshold. One use everybody data together, the other use each group data separate. I wan see which groups dem dey refer to person more often when the model no sure.

Read the study Paper before peer review

Gap for coverage / Gap for human review

One threshold for everybodySeparate threshold for each group

Gap for coverage

5.0
1.2

Gap for human review

5.1
12.9
Percentage pointsTable 4
Some of my work

Research wey I don publish

  1. Model

    Turkish Sentiment Analysis · TSAM

    I train dis BERT model to pick di tone wey Turkish review get. E dey give separate score for positive, neutral and negative tone.

    2022
  2. Paper before peer review

    When coverage and human review no balance

    Whose case dem go check again?

    2025
  3. Code repository

    Agent Chronicle

    The coding agent say e don finish. Wetin e change?

    2026
  4. Journal article

    How sure quantum predictions fit be

    How much a few measurements fit tell us?

    2025
  5. Paper before peer review

    TARA and quantum measurements

    When a stream of measurements begin look unusual?

    2025
  6. Patent application

    Assistant wey fit answer and act

    When chat suppose turn to action?

    2025
  7. Patent application

    Small robot for rescue work

    Wetin fit reach person through space wey wide like one finger?

    2025
  8. Patent application

    Talk to a POS machine

    Shop owner fit control payment terminal with voice?

    2024
  9. Patent application

    Keyboard and mouse for virtual space

    If the screen dey virtual, where the keyboard dey?

    2025
  10. Patent application

    Advice wey suit each worker

    How workplace assistant fit choose advice for one person?

    2025
  11. Paper before peer review

    Put quantum test results to the test

    Good score go still hold if conditions change?

    2025
  12. Journal article

    From Turkish speech to writing

    Wetin model wey don hear many languages fit use when e learn Turkish?

    2024
  13. Model

    Small Qwen model for maths

    Small model fit learn how to arrange a maths solution?

    2025
  14. Dataset

    TARA tests for reasoning in Turkish

    How we fit test whether model dey reason well in Turkish?

    2025
  15. Journal article

    One language model for many chatbots

    Thirteen chatbot tasks need thirteen separate models?

    2021
  16. Journal article

    When advert get wrong label

    Wetin if na the label on the advert dey wrong?

    2021
  17. Code repository

    How language models fit edit neural networks

    One model fit change another model weights?

    2026
  18. Journal article

    BLOOM

    How researchers build one multilingual model for everybody to study?

    2022