I train neural networks from scratch, mostly in JAX / Flax. Two lines of work:

• Language models — the Julian family: bilingual (EN / FR) LLMs, from raw data to instruction-tuned chat, on TPU.
• World models — oneiro: a from-scratch DreamerV3 agent that learns its entire policy inside the imagined rollouts of its own world model.

_{oneiro (~64k env steps) unlocking achievements in Crafter — learned entirely from imagination rollouts}

A from-scratch reimplementation of DreamerV3 (Hafner et al., 2023) in JAX / Flax NNX, trained on the Crafter benchmark (sparse rewards, 64×64 pixels, 22 hierarchical achievements).

• 15.26M parameters — DreamerV3-S class
• Rainbow-level in 64k env steps: 4.0 achievements/episode, vs Rainbow's 4.3 at 1M steps — a ~16× sample-efficiency gap
• RSSM + imagination-based actor-critic, GPU-resident replay buffer, lax.scan everywhere
• 18+ documented training runs with full debugging journals and a hypothesis registry — the real value of the repo

→ JulianKerignard/oneiro

Wikipedia EN/FR dumps  →  clean / tokenize (SentencePiece)  →
pretrain (JAX · Flax · Optax · FSDP on TPU)  →  SFT / instruct (ChatML)  →  ship on the Hub

Plus a write-up of the whole thing: julian-600m-paper.

Datasets I built & published: wikipedia-en-julian  ·  wikipedia-fr-julian — the bilingual pretraining corpus.

• Model Context Protocol servers — Blender · Unity · Obsidian · Trello · webcheck
• ClawdEngine_Rust — a 3D engine in Rust  ·  StatsMonitor_MacOS — Swift