Skill
4 entries mention it. All skills →
Studied how small models learn when given reasoning scaffolds, and compared the insides of base and fine-tuned models.
GRPO, LoRA, MLX, Mechanistic interpretability, Experiment design
Built command-line tools that show a model’s predictions layer by layer, compare activations between models, measure drift and run ablations.
MLX, Activation analysis, Logit lens, Model comparison, CLI/TUI, Mechanistic interpretability
Developed training tooling for Apple Silicon, with configurable tasks, evaluators, rewards, supervised modes and guided reasoning experiments.
MLX, Python, GRPO, LoRA, SFT, Reward design, Training infrastructure
Kept training small models at home: guided-reasoning runs, quantisation and tooling to make runs repeatable.
MLX, GRPO, LoRA, Quantization, Dataset engineering, Experiment tracking