Repository navigation
Preserve refusal content in SparkGLM Chat Completions - #3
Merged
data-angel merged 1 commit intoOct 13, 2026
Merged
Conversation
Keep the full blocking assistant answer when the heuristic refusal classifier matches, consistent with streaming. Retain refusal annotations, tool-call gating, and finish reasons. Add production assembly/serialization and streaming regressions; update the compatibility expectation. Reported in Atlas-Inf#172.
|
Thanks for the contribution. Before we can merge, please sign our Contributor License Agreement by replying to this comment with:
I have read the CLA Document and I hereby sign the CLA You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot. |
volfco
pushed a commit
to volfco/atlas
that referenced
this pull request
Oct 17, 2026
…rsion PRs off Two maintainer decisions from the 2026-09-27 PR triage, both aimed at repo and PR-queue bloat: - CONTRIBUTING: an opt-in switch whose A/B is neutral or worse records its finding in the [job] issue instead of merging (e.g. Atlas-Inf#97's bf16a MoE arm). - dependabot: open-pull-requests-limit 0 for cargo and actions version updates. Six stale bump PRs (Enntity#3-Atlas-Inf#6, Atlas-Inf#49, Atlas-Inf#50) were closed; security updates still come through, and bumps get taken in one deliberate batch. 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Preserve full blocking Chat Completions content when the refusal-prefix classifier matches, alongside the existing refusal metadata. This fixes the empty assistant body reported in Atlas-Inf#172. Streaming already preserves the text; its regression coverage is extended here. Classifier behavior and tool gating remain unchanged.
This is a one-commit maintenance update on the published SparkGLM RC4.2 engine. Responses and Anthropic adapters receive the preserved text through their existing conversion paths; their refusal-metadata contracts are unchanged.
Validation: 25 refusal tests, 43 streaming tests, formatting and diff checks passed. The new content tests reproduce four failures before the fix. Independent review found no material blocker. Native arm64 CUDA image build passed all configured gates, including 25 refusal and 43 streaming tests. Published image:
ghcr.io/enntity/atlas-sparkglm@sha256:5ec57d5522a95463b600a17471ce60712b315566e04c0a3bb6af1d84a1de0a4a. Live model activation awaits a coordinated window on the pair, which is reserved for MTP experiments.Local whole-workspace Clippy cannot pass on this Mac: Linux-only IoUring test references and newer Rust 1.96 lints in unchanged plugin/model files fail before this patch. The image uses the pinned Rust 1.93.1 toolchain.
Fork CI limitations, inspected on this PR: file-cap and typo failures are confined to unchanged files (every changed Rust file is <=333 lines). Workspace tests, coverage, and benchmark linking fail on the unchanged CI cuBLASLt stub missing
cublasLtMatrixLayoutSetAttribute. The kernel job failed downloading serde from crates.io. The inherited CLA workflow cannot find the fork'scla-signaturesbranch. These checks are not green. Pinned-toolchain Clippy, formatting, SPDX, security audit, ancestry, and provenance checks passed; the native GB10 release-image build and configured regression gates passed.