Configuration Guidance

Practical guidance for composing interpretune configurations. See Session, Module, and DataModule Usage Patterns for the session/config API itself.

Keep the composition space flat

Interpretune deliberately keeps configuration composition flat: configs compose by explicit inclusion/override rather than deep implicit inheritance hierarchies. When authoring new configs, prefer adding a sibling config (or a small override file) over introducing new nesting levels — flatness is what keeps the composition space legible and greppable as the config surface grows.

Group configs by model, then experiment

Configuration files can technically be composed arbitrarily, but the convention used throughout the examples — and recommended for your own experiments — is grouping by model level (shared per-model substrate settings) and experiment level (task/experiment-specific settings layered on a model config). See src/it_examples/config/experiments/ for the reference layout (e.g. rte_boolq/gemma2/..., rte_boolq/gemma3/...).

HookedTransformer: model_name_or_path can be arbitrary

When instantiating a TransformerLens HookedTransformer via interpretune config, the ITLens config’s model_name_or_path can be set to an arbitrary identifier — the TL pretrained-model resolution is driven by the TL config’s own model selection, so model_name_or_path functions as a label rather than a required HF-hub pointer in that mode.

CLI: auto_comp_cfg is not supported via jsonargparse

The auto_comp_cfg mechanism (automatic composition-config resolution, interpretune.config.shared.AutoCompConfig) is not supported through the CLI: jsonargparse’s signature-driven parsing cannot represent its dynamic target-class rebinding. Configure auto_comp_cfg programmatically (or via YAML consumed by the session API) instead of CLI flags.

Lightning precision plugins only convert floating-point tensors

When composing with the Lightning adapter, note that Lightning precision plugins convert inputs via apply_to_collection(..., function=...) targeting floating-point tensors only. Integer tensors — token IDs, attention masks, embedding indices (LongTensor) — pass through unconverted, which is exactly what embedding layers require. If you see mixed-precision surprises, audit the float tensors in your batch; the integer inputs are not the issue.