Conversation
Gemma leaks <|channel>thought, <think>, and a leading Thought. into content. Strip those before TTS so they are never spoken. Co-authored-by: Neevash Ramdial (Nash) <mail@neevash.com>
Interrupted replies still record tools in history. Barge-in runs them; a changed ask injects cancelled results. Restaurant prompt books and orders in the same turn once the lookup succeeds. Co-authored-by: Neevash Ramdial (Nash) <mail@neevash.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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
Restaurant bench runs were failing because Gemma spoke thought-channel markup (
Thought.,Channel thought) and interrupted replies dropped in-flight tool calls.Changes
<|channel>thought,<think>, and a leading leakedthought/Thought.from OpenAI-compat content before it is spoken.check_availability/lookup_menusucceed, callcreate_reservation/create_orderin the same turn when name+allergen / items+allergen are known.Gemma k=1 validation
Run
20260909T014745Z, restaurant pack, Flux STT,gemma/gemma-4-26B-A4B-it.Pass 3/8. adversarial / coherence / tool_filler pass; entity_dense, golden, interrupt, noise_kitchen, selectivity fail.
V2V P50 1520 ms (n=22). False cutoffs 1.00/call.
Thought-channel leakage: gone from speech (no
Thought./Channel thoughtin transcripts). One rare router log leftovertext=<channel|>.Log:
/tmp/va-bench-gemma-after.log. Writeup:benchmark/out/gemma-k1-after.md.