CHANGELOG.md

Changelog

v1.3.0

Breaking Changes

• Python 3.10+ required — dropped Python 3.9 support.
• nnsight >=0.6 required — bumped minimum dependency from 0.5 to 0.6.
• Replaced loguru with standard library logging — all logging now uses logging.getLogger("nnterp"). Users who configured loguru sinks for nnterp should switch to logging.getLogger("nnterp").setLevel(...) etc.

New Features

• token_positions and batch_index parameters for steer() — fine-grained control over which tokens and batch elements are steered. Both can be combined:

  with model.trace(["prompt A", "prompt B"]):
      model.steer(layers=1, steering_vector=v, batch_index=0, token_positions=[0, 1])

  The old positions parameter is deprecated (still works, emits DeprecationWarning).

• remote parameter on StandardizedTransformer — remote=True automatically sets allow_dispatch=False and registers nnterp for NDIF remote execution.

• Multimodal model detection — check_model_renaming() now detects heterogeneous layer types (e.g. self-attention + cross-attention in vision-language models) and raises an informative error. Bypass with allow_multimodal=True.

Internal

• Removed broken steer import from nnterp.interventions in tests/demos (was already a dead reference).
• Added comprehensive steer tests covering token_positions, batch_index, and their combination.
• Updated README with new branding, corrected documentation URLs, and new steer API examples.