Imported from geminiappgabdemouser404-cmd/applied-ml (Repos/mechlens/transformer_lens/config/AGENTS.md). Install upstream with npx skills add geminiappgabdemouser404-cmd/applied-ml --skill config. Copyright stays with the author.

Config - AGENTS.md

The config dataclasses that drive both HookedTransformer and TransformerBridge. Read the root AGENTS.md for project-wide rules.

File map

Adding a new HF-config attr to TransformerBridgeConfig - decision tree

The logit_scale bug existed because the rules below weren't documented anywhere. Pick exactly one of the four paths; doing more than one creates silent-override risk.

Don't declare the same attr both as a dataclass field AND a PASSTHROUGH entry - PASSTHROUGH writes happen AFTER from_dict, so the runtime value will silently overwrite whatever defaults / explicit handlers set.

Existing properties that may surprise you

• head_dim is a read-only @property aliasing d_head (line 212). setattr(cfg, "head_dim", val) raises AttributeError: property 'head_dim' has no setter. Don't add it to PASSTHROUGH.
• n_heads has = -1 as its placeholder in the constructor signature, deliberately. Comment in the source: "Add n_heads to signature so it's not filtered out by from_dict". Don't "fix" this default.

Verifying a new field propagates

Integration test pattern (the regression test that would have caught the logit_scale bug):

def test_cfg_<attr>_matches_hf(bridge: TransformerBridge, hf_model: Any) -> None:
    """Regression: <attr> must propagate from HF (not silently fall back)."""
    bridge_val = getattr(bridge.cfg, "<attr>")
    assert bridge_val == hf_model.config.<attr>

The getattr form sidesteps mypy's "undeclared field" complaint without # type: ignore (see AGENTS.md §10).

Pick a test model whose <attr> value differs from the adapter's hardcoded fallback (if any) - otherwise the assertion passes by tautology.