Serkan Altuntaş

Software Architect in Istanbul, with a computational biology background. Learning LLM systems in public through open-model evals, small-model training, and reproducible engineering notes.

Latest in Journal

GCTX: the first full model training and what followed

Oct 13, 2026 · note · GCTX completed its first full 93M-parameter training run on September 22. Follow-up experiments show reliable small-set fitting, while useful generalization remains open.

Why I Froze an Imperfect Eval Baseline

Jul 31, 2026 · note · A follow-up note on why the July Open Model Lab baseline stays frozen even after the failure audit found grader artifacts, strict-format misses, and token-limit effects.

Open Model Research Harness completes its July evaluation gate

Jul 29, 2026 · note · The first Open Model Research Harness gate is complete: 25 tasks, three local open models, 75 scored outputs, a deterministic comparison, and a failure audit that shows why raw pass rates need context.

Starting Open Model Research Harness

Jul 4, 2026 · note · A public 12-stage project to build an eval-first research-engineering harness for open LLMs.

Local model memory: bits, dense, MoE

Jul 2, 2026 · note · A short note for reading local model memory requirements without confusing active MoE parameters with loaded weights.

GCTX passes its first data-readiness gate

Jun 28, 2026 · note · GCTX's data pipeline passed its first volume, split, and format checks for a proof-model run.

All journal notes →

Featured Projects
Open Model Research Harness Building

A 12-stage public research-engineering project for reproducible open-model evals, post-training, agents, safety, monitorability, and systems efficiency.

Fresh 8% Next: SFT Pipeline + Data Quality
GCTX — a language model for Git diffs Building

A from-scratch language model for Git diffs and Conventional Commit messages. The first 93M-parameter training run is complete; useful generalization remains open.

Fresh 54% Next: Prepare 512 examples checked against their source diffs
AI Usage Standard (AIUS) Alpha

AI Usage Standard (AIUS): a deterministic disclosure system for published artifacts, with six dimensions and validated frontmatter.

Idle 54% Next: Phase 2 — npm package (@serkanaltuntas/ai-usage-badge)

All projects →

About

Fifteen years of software, with a computational biology and bioinformatics background. Currently working through LLM systems in public: open-model evals, small-model data pipelines, post-training experiments, and reproducible reports with explicit claim boundaries.

Full bio →