Kacper Wikiel

I build pretraining data, tokenizers, and evaluation pipelines for Polish and low-resource language models.

My current work is the Polish LLM stack around SlayerLab: a 6.28B-token open Polish corpus, custom byte-level BPE tokenizers, runnable GPT-style checkpoints, replication docs, and benchmarks where the numbers come from running the models myself.

I have taken this work about as far as one person with local compute can take it. Four RTX 3090s under my desk are enough to test ideas, expose bad assumptions, and build reproducible artifacts. The next step is doing the same work inside a real lab, with real compute and colleagues.

I came up through physics, so I would rather measure a thing than trust a demo of it. That instinct shows up in how I build datasets, choose model baselines, inspect tokenizer failures, and decide whether a benchmark is measuring anything real.

LLM work

Polish DynaWord - 2.49M documents and 6.28B tokens of Polish text. I curate the corpus, review source licenses and provenance, keep the build reproducible, and document the training-mix problems instead of hiding them in a dump.
Slayer GPT Tokenizer Model - a runnable GPT-style Polish checkpoint with its tokenizer and training artifacts. The repository includes custom byte-level BPE tokenizers, model code, logs, replication notes, and the compatibility rules needed to reproduce the experiment correctly.
Continued-pretraining pipeline for Polish LLM work. I work on data mixture design, base-model tradeoffs, tokenizer evaluation, corpus preparation, and evaluation on Polish tasks such as MTEB-style retrieval and public-sector extraction benchmarks.

Evaluation

CodeSOTA - a price-aware benchmark site for machine-learning models. Around thirty thousand people read it every month. I care about benchmarks where the methodology is explicit, the models are actually run, and cost is part of the result.
ZusWaveBench - NER evaluation on Polish government documents. I built a unified pipeline across 46 LLMs with strict and relaxed metrics, price-quality recommendations, and failure analysis for real Polish administrative text.
Tokenizer and corpus evaluation. I look at source-level perplexity, style contamination, decontamination, vocabulary size, Polish morphology failures, and what happens when a corpus is legally clean but distributionally wrong.

Things I'm building

SlayerLab - open Polish pretraining data, tokenizer experiments, and reproducible small-LM artifacts.
CodeSOTA - model evaluation where the numbers come from running the models myself.
Hardparse - an OCR API that reads tables, formulas, charts, and seals.
Bozenka - an offline Polish text-to-speech voice.

Applied systems

Computer vision that reads the welds running the length of a national fuel-pipeline network. Building it, I turned up a sensor bias the earlier analysis had walked straight past.
A Raspberry Pi that finds aircraft, follows them with a gimbal, and fixes their GPS position from one camera in real time. Most of the AI worth having is going to run on small hardware like this.
A document platform with attribute-based access control and a hash-chain audit trail, for work where you have to prove where every record came from.

Experience

SlayerLab Pretraining data, tokenizers, evaluation2026
PwC ML Engineer, contract2025
National Pipeline Operator ML / CV Engineer2024–25
Procter & Gamble Software Engineer, Generative AI2022–24
Igoria Trade Python Developer2021–22

Media

TVN24, Czas Przyszly: "Media spolecznosciowe dla botow AI". I talked with Wojciech Skrzypek about AI-only social spaces, Claude-like agents, and what changes when agents start producing and reacting to online content.
Wycinki z Przyszlosci. I talked about Slayer, CodeSOTA, why Poland might want its own language model, compute as the real bottleneck, benchmarks, open-source artifacts, and the path from lurker to contributor.