Coccinella Labs

Coccinella Labs builds small, inspectable models that run on consumer hardware, and publishes the training code alongside the weights.

What we work on

Fine-tuning and small language models. DistilBERT and BERT encoders adapted for extractive question answering and named entity recognition, and GPT-2 language models, both fine-tuned and trained from random initialisation.

Speech recognition. Whisper checkpoints adapted for short spoken passages, with the tokenizer and feature extractor published alongside the weights so the repository loads on its own.

Apple Silicon. MLX conversions of instruction-tuned language models, quantised to four bits for machines without a discrete GPU.

Reinforcement learning. CMA-ES policies for classic control tasks, with the environment metadata and weights stored so a published checkpoint can be loaded and evaluated directly.

System telemetry. A macOS metrics capture, published as a dataset for time-series and anomaly-detection work.

Areas

fine-tuning · transformers · apple-silicon · mlx · speech-recognition · reinforcement-learning · small-models

Tasks

question-answering · token-classification · text-generation · automatic-speech-recognition · reinforcement-learning

Libraries

transformers · mlx · safetensors · gymnasium · pytorch · pandas

Base models

distilbert-base-uncased · bert-base-uncased · gpt2 · whisper-small · SmolLM-135M-Instruct-4bit

Where the models live

The published models and datasets are under the harpertoken namespace on the Hub, and the training and serving code is on GitHub.

Every model card states what was measured and what was not. Where an earlier card quoted scores that no evaluation supports, those numbers were removed rather than restated.

Contact

coccinella.labs@icloud.com