2023–26Independent and Invisible TechnologiesHugging Face · GitHub · frontier labs via Invisible
Training, merging and porting models
Before I wired models into enterprise stacks I trained them. A year of human-feedback work on frontier LLMs, 46 open models and 7 datasets of my own, and a port of a new brain-inspired architecture to Apple silicon.
- Outcome
- 188k downloads of models I trained, re-quantized by TheBloke, bartowski and others
- Role
- Researcher and engineer, from dataset to weights to quantized release
- Stack
- PyTorch · Transformers · MLX · Unsloth · LoRA · MergeKit · llama.cpp · GGUF · GPTQ · AWQ · Jamba · Liquid LFM
- Links
- Hugging Face BDH-MLX Vodalus
- models released
- 46
- models released
- all-time downloads
- 188k
- all-time downloads
- training throughput after the MLX port
- ~2×
- training throughput after the MLX port
Training with human feedback
At Invisible Technologies I spent a year on the training side of frontier language models. I reviewed model outputs, named the behavioral patterns behind failures and wrote the natural-language feedback that steers them. I prepared data for both general-purpose and specialist models and wrote the quality rubrics the projects scored against. It is the most useful thing I know when a customer asks why a model behaves the way it does.
Open models
On my own time I fine-tune, merge and release. ANIMA was an early biomimicry language model: a Mistral 7B fine-tune on biomimicry, design-process and STEM data, downloaded 67k times. Nexus and Einstein were trained on Internal-Knowledge-Map, a dataset I built to teach models to connect what they already know. I release LoRA adapters and GGUF, MLX, GPTQ and AWQ builds so people can run them on the hardware they own, and I merged mixture-of-experts models and trained early Jamba hybrids of state-space and attention layers when they appeared.
- Models
- 46
- Datasets
- 7
- Downloads
- 187,876
- On community builds
- 137,519
Base modelMy fine-tune or mergeMy buildCommunity build
A new architecture on a laptop
In September 2025 Pathway published the Dragon Hatchling (BDH), a language model built more like a brain than a transformer. I ported it from PyTorch to Apple MLX so it trains in unified memory, then built mixture-of-experts and recursive-depth variants, a custom BPE tokenizer and synthetic and puzzle training runs to see what it can learn.
The port roughly doubles throughput on an M2 Max: about 5,000 tokens a second against 2,500 on PyTorch MPS, in 8 GB of memory instead of 12. The curves below are from my own training logs.
Recursive variant, 32k BPE tokenizer, PleIAs SYNTH
Validation loss 4.45 → 3.66 · steps 20.1k–31.5k
Byte-level BDH, 256-symbol vocabulary, PleIAs SYNTH
Validation loss 2.61 → 2.35 · steps 100–10k
Tooling
Vodalus, 195 stars on GitHub, builds fine-tuning datasets grounded in RAG and Wikipedia and trains them efficiently with MLX and Unsloth. It is the pipeline behind several of the models above.