dir-steering: pass --think, not --think-high, to ds4 - #935
Open
eauchs wants to merge 1 commit into
Open
Conversation
build_direction.py --think aborts before capturing anything:
ds4: unknown option: --think-high
The ds4 CLI accepts --think, --think-max and --nothink; --think-high is
not one of them, so every thinking-mode capture fails at the first pair.
The non-thinking path is unaffected, which is why the default run works.
Verified on macOS with GLM 5.3 Flash Q2: with the fix, --think completes
100 pairs and writes a 45 x 4096 direction, and the resulting direction
is nearly orthogonal to the one captured on direct answers (mean per-layer
cosine 0.19), so the two capture points are not interchangeable.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
build_direction.py --thinkaborts before capturing anything:ds4accepts--think,--think-maxand--nothink.--think-highis not one of them, so every thinking-mode capture fails on the first pair. The default (non-thinking) path is unaffected, which is why the documented example still works.One character in
run_capture():Verified on an M3 Max 128 GB, GLM 5.3 Flash Q2,
--profile glm-5.3-flash:Worth having because the two capture points are not interchangeable. Building the same 100 matched pairs both ways, the per-layer cosine between the direct-answer direction and the post-
<think>direction averages 0.19, diverging most in layers 3 to 12. So--thinkselects a genuinely different behavioural axis rather than a variant of the same one — and until now it was unreachable.