OLMo2-1B EEG Code Generator (v8)
A 1.3B parameter language model fine-tuned for EEG/LFP neuroscience code generation,
specializing in Parkinson's disease research.
Model Description
Built at the Neural Dynamics and Modulation Lab (NDML), Cleveland Clinic.
Trained using OLMo-Core on a mixed web + code corpus, then fine-tuned on
synthetic EEG analysis examples covering MNE-Python, EEGlab, FieldTrip, and Chronux.
Base model: OLMo2-1B (mixed web + Python + MATLAB pretraining)
Fine-tuning: 8 iterations on EEG toolbox code generation pairs
Best benchmark score: 0.937 (30/30 prompts passing)
Capabilities
- MNE-Python EEG preprocessing and analysis
- Beta band oscillation analysis for Parkinson's disease
- LFP analysis from DBS lead recordings
- FieldTrip, EEGlab, and Chronux MATLAB toolbox code
- GPi/STN local field potential analysis
- Closed-loop DBS (eiDBS) analysis pipelines
Usage
Training Details
- Parameters: 1.3B
- Pretraining: 868M web tokens (Dolma v1.7) + 654M mixed web/code tokens
- Fine-tuning data: ~3,200 synthetic Claude-generated EEG code pairs
- Hardware: NVIDIA A100 80GB PCIe
- Framework: OLMo-Core v1.9.0
Lab
Neural Dynamics and Modulation Laboratory
Cleveland Clinic Lerner Research Institute
PI: David Escobar Sanabria, PhD
