🧠 Building Your First Encoding Model

Learn the fundamentals of neural encoding models with LITCoder

1. Choose Your Parameters

Start with these basic settings for your first model:

Dataset

Narratives (easiest to start with)

Model

GPT-2 small (fast and reliable)

Layer

Layer 9 (good middle layer)

Downsampling

Lanczos (recommended)

2. Basic Code Example

Here's a simple example of how to build an encoding model:

# Create assembly for your dataset
assembly = AssemblyGenerator.generate_assembly(
    dataset_type="narratives",
    data_dir="data/narratives/neural_data",
    subject="sub-256",
    tr=1.5,
    lookback=256,
    context_type="fullcontext"
)

# Extract features from language model
extractor = LanguageModelFeatureExtractor({
    "model_name": "gpt2-small",
    "layer_idx": 9
})

# Downsample to brain timing
downsampler = Downsampler()
features = downsampler.downsample(
    data=model_features,
    method="lanczos"
)

# Apply FIR delays
delayed_features = FIR.make_delayed(features, delays=range(1, 9))

# Train the model
metrics = fit_nested_cv(
    features=delayed_features,
    targets=brain_data,
    n_outer_folds=5
)

3. What Happens Next

After running your model, you'll get:

Correlation scores - How well your model predicts brain activity
R² values - Proportion of variance explained

4. Next Steps

Once you've built your first model: