Skip to content

Pre-release validation: executed notebook outputs + validation fixes (post #2157)#2187

Draft
romanlutz wants to merge 10 commits into
microsoft:mainfrom
romanlutz:romanlutz-validate-pr-2157
Draft

Pre-release validation: executed notebook outputs + validation fixes (post #2157)#2187
romanlutz wants to merge 10 commits into
microsoft:mainfrom
romanlutz:romanlutz-validate-pr-2157

Conversation

@romanlutz

Copy link
Copy Markdown
Contributor

Purpose

This is a draft validation/review branch, not a clean mergeable change. It exists so the executed notebook outputs and validation repairs can be reviewed in a diff. It contains committed notebook outputs and several Merge origin/main commits (latest main: #2178 FlipAttack→flip technique, #2158 harm-category standardization), so it is not intended to be merged as-is.

What's in here

  1. Executed notebook outputs committed across the doc corpus so the rendered results are reviewable (this is the main thing to eyeball).
  2. Deterministic validation repairs surfaced during exhaustive post-merge validation:
    • Typed scenario DTO serialization, renamed adversarial first-message handling, per-objective resume scheduling, benchmark/adversarial target selection.
    • Scanner backend-identity health check + isolated --start-server port; Windows CP1252 encoding-safe output paths.
    • Hugging Face snapshot_download + direct local load (failures no longer swallowed).
    • Reference-counted Azure async token provider lifecycle; OpenAI/realtime targets and the concrete AzureContentFilterScorer close clients/credentials idempotently (no generic scorer-cleanup contract added, per review).
    • HuggingFace unit test-fixture coroutine leak fixed.
  3. Latest origin/main merged (MAINT: Migrate FlipAttack to a core 'flip' attack technique #2178, FEAT Overlay harm-category standardization across dataset loaders #2158, and prior) with the two conflicted notebooks regenerated from their merged .py sources.

Validation evidence (high level)

  • Unit: 10,629 passed / 6 skipped with uv sync --all-extras; affected post-merge groups 110 (MAINT: Migrate FlipAttack to a core 'flip' attack technique #2178) + 2,266 (FEAT Overlay harm-category standardization across dataset loaders #2158 datasets/models) passed.
  • Notebooks: authoritative corpus 64/64 executable notebooks passed; 271 output cells / ~730–795 output objects manually audited; 0 error outputs. 8 notebooks remain external/manual (SQL, quota, Azure ML, Copilot infra).
  • Integration: 155 pass after fixes; remaining 41 fail / 8 error are external (credentials, quota, SQL DNS, missing services).
  • Python E2E: repaired regressions verified (Leakage 5/5, RapidResponse 2/2, AdversarialBenchmark 7/7).
  • Frontend: Jest 833/833; lint/type/build pass; Playwright mock 61 (+1 intentional skip), seeded 61, live 24/24.
  • Pre-commit: all hooks pass on the full tree, including ty with all extras.

Full evidence-based report with per-notebook/per-cell questionable outputs is in the session artifacts (files/validation-report.md).

Draft on purpose. Please review the executed notebook outputs; do not merge this branch to main as-is.

Co-authored-by: Copilot App 223556219+Copilot@users.noreply.github.com
Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448

Copilot AI added 10 commits July 12, 2026 14:36
Keep scenario DTOs typed for CLI output, migrate the custom-target example to first_message, and remove empty per-objective attacks when reconstructing sampled scenario resumes.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Repair Python E2E target propagation and notebook examples, correct the
Responses vision payload, and isolate and harden frontend E2E validation.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Repair scanner isolation, provider resource cleanup, Hugging Face downloads, Windows-safe output, and executable notebook examples discovered by exhaustive post-merge validation.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Use the newly configured adversarial_chat target in the benchmark end-to-end smoke test after confirming registration and a clean 7/7 full scenario run.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Remove the generic scorer lifecycle method and threshold-wrapper delegation while retaining Azure Content Safety's explicit resource cleanup.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Store the audited execution results for all runnable documentation notebooks so output cells can be reviewed directly in the branch changeset.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Retain Realtime WebSocket cleanup as a public API while moving ordinary OpenAI teardown behind a private hook and removing validation-only cleanup calls from documentation notebooks.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
…lidation branch

Resolved notebook conflicts in doc/code/converters/1_text_to_text_converters and doc/code/executor/1_single_turn by regenerating executed outputs from the merged .py sources. Verified affected tests (110 passed), full pre-commit, and notebook cell-output audit (0 errors).

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: 52d2518d-ebe5-4888-8189-6146b6503448
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants