[Proposal] Rewrite executable doctests that construct HookedTransformer onto TransformerBridge

[jlarson4]

Proposal

Base on dev-4.x. Rewrite the executable >>> doctests that build a HookedTransformer so they construct a TransformerBridge instead, in the modules that keep running doctests after HookedTransformer is removed: transformer_lens/ActivationCache.py (4 HT-constructing doctests), transformer_lens/evals.py (2 HT-constructing doctests, only 1 executable at :326-329 — the other is already # doctest: +SKIP at :554-559), transformer_lens/utilities/exploratory_utils.py (1).

Motivation

HookedTransformer is being deprecated in favour of TransformerBridge. ActivationCache, evals, and exploratory_utils are staying, but their doctests still do HookedTransformer.from_pretrained(...). pytest runs these (pyproject.toml enables --doctest-modules over transformer_lens/), so they will hard-fail the moment HookedTransformer is removed. Migrating them now keeps the examples current and unblocks removal.

Pitch

• Replace HookedTransformer.from_pretrained("gpt2"...) with TransformerBridge.boot_transformers("gpt2") (add .enable_compatibility_mode() where the example relies on HookedTransformer-exact hook names / numerics).
• Update expected doctest output to whatever the bridge prints, and any # doctest: +SKIP markers that are no longer needed. Note every executable example also asserts the literal output line Loaded pretrained model ... into HookedTransformer, so each rewrite must update that expected-output line too.
• Keep each example's teaching intent identical; only the model construction changes.

Acceptance:

• No doctest in ActivationCache.py, evals.py, utilities/exploratory_utils.py constructs HookedTransformer
• make docstring-test (or pytest --doctest-modules transformer_lens/) passes
• make unit-test passes
• uv run mypy . passes

Mirror: existing bridge examples in docs/source/content/ and the boot_transformers / enable_compatibility_mode usage in tests/unit/model_bridge/.

Checklist

• I have checked that there is no similar issue in the repo (required)

[Canonik]

Hi @jlarson4, I’d like to take this on based on dev-4.x. I’ve previously contributed model architecture support and tests to TransformerLens, so I’m already familiar with the repository’s testing conventions.

[jlarson4]

For sure @Canonik ! You are now assigned