Repository navigation
Fix issues from repo review - #4
Merged
Merged
Conversation
…g, logs, packaging - Degeneration fallback re-encodes the prompt after dropping vectors, so the unsteered retry no longer tells the model they are still active. - apply()/drop() called between steps now match the model's own tool calls: one full-strength steered turn, and a COOLDOWN_TURNS-long cooldown. - A harness's leading system message is merged into the steering prompt instead of producing two system messages. - steer rejects non-finite strengths (inf), not just NaN. - Raise on zero/non-finite vector norms instead of steering with NaN. - torch.load uses map_location="cpu"; extraction always removes its hook. - Demo logs: record the --seed prompt, log inputs under the turn they feed, and rename the "decay" event to "steering_state" (replay reads both). - Add pyproject.toml (pip install -e .) and drop unused config constants. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… setup Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #3 (rename); fixes from the internal repo review.
Fixes
apply()/drop()between steps were off by one: a manually applied vector got two full-strength turns, and a manual drop blocked re-use for three turns. Both now match the model's own tool calls.systemmessage is merged into the steering prompt instead of sending two.steernow rejectsinf(previously only NaN); zero or non-finite vector norms raise instead of injecting NaN.torch.load(..., map_location="cpu"); extraction always removes its capture hook.--seedprompt is now logged; inputs are logged under the turn they feed into; the misleading"decay"event is now"steering_state"(replay.pystill reads old logs).pyproject.toml(pip install -e .) and removed unused config constants. AGENTS.md documents both, and notes that Gemma-2 (nosystem/toolroles) isn't supported yet.Deliberately not changed:
reportfeedback still isn't written toregistry.yaml. It's available inStepResult.tool_resultsand the session logs; where it should be persisted is a design call.Testing (RTX 4090 RunPod, Qwen2.5-7B-Instruct)
pytest: 32 passed (CPU + GPU). 6 new CPU tests, all confirmed to fail on the old code.verify_vector.py --vector assistant_axis --strength 12: clear ± shifts versus the baseline../run.sh --auto --max-turns 5: the model steered itself to -15 and degenerated; the detector dropped the vector and the retry was coherent. The replay shows the seed and the new turn labels.🤖 Generated with Claude Code