AI.004 · Planned
Specialising a sub-1B language model for Pixel, using rented GPU compute for fine-tuning and the Intel N100 as the real deployment target. The experiment focuses on conversation, tool use and useful proactivity rather than general model capability.
Research brief
Can a sub-1B model be fine-tuned specifically for Pixel to improve conversation, tool use and proactive behaviour while remaining practical to run locally on an Intel N100?
Pixel does not need a small model that knows everything. It needs one that understands its role, its tools and the environment around Afterlab.
A specialised model could trade broad general knowledge for better behaviour on the tasks that actually matter — while keeping routine interactions local and reserving larger models for harder reasoning.
Training
Rented GPU compute.
Deployment
FIREBAT AK2 Plus, Intel N100, 16 GB RAM.
Model size
Sub-1B, with the final base model frozen before training.
Runtime constraint
The resulting model must be small and responsive enough to remain locally available for Pixel.
Architecture constraint
The model selects, proposes and responds; the Agent remains responsible for tool execution, permissions and protected operations.
Evaluation constraint
The benchmark must be frozen before fine-tuning so the base and specialised models are tested on the same cases.
A small model fine-tuned on Pixel-specific behaviour will outperform its base version on tool use and operational context without losing conversational quality, while remaining fast enough for local inference on the N100.
Keep exploring
Every experiment starts with a question and ends with notes worth keeping. Browse the rest of the lab, or see which ideas grew into full projects.