experiment(training): evaluate Jev-Omni as Studio pre-review sidecar - #30
Merged
Merged
Conversation
Adds an offline, replayable experiment (training/jev + run_jev_experiment.py) that materializes residual Studio review rows into digest-verified records, routes them through the local Jev-Omni-MLX-4bit worker, and fits calibrated routing thresholds on a source-grouped holdout. Measured on 929 residual rows across 35/70/140 image-token budgets and two prompt variants: auto-decisions reach 100% accuracy at safe thresholds but reduce review by at most 9.3%, far below the 25% gate, and auto-reject never engages because most reject reasons are contextual rather than visible in the crop. Recommendation: remove. Fixes #25 Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Implements the offline, replayable Jev-Omni evaluation required by #25:
training/jev/materializes residual Studio review rows (rows the deterministic auto-accept/auto-reject rules do not decide) into digest-verifiedPreReviewRecords: crop SHA-256, de-duplicated OCR candidates, bounded options (+ "none correct" / "not a valid target"), engine evidence, and human truth — ground truth excluded from the input digest.OmniMlxRunnerrunsRuiruiz30/Jev-Omni-MLX-4bitlocally in an isolated stdlib worker subprocess inside its own venv — nomlx/mlx-vlmin project dependencies, no external service, and production/training paths untouched.decisions.jsonl;--replayreproduces reports without model calls.Measured result (929 residual rows, 35/70/140 image-token budgets, two prompt variants): auto-decisions are 100% accurate at safe thresholds but reduce review by at most 9.3% — far under the gate — and auto-reject never engages because most reject reasons are contextual rather than visible in the crop. Report recommendation: remove; final lifecycle call stays with the maintainer per the issue.
training/README.mddocuments reproduction and the measured outcome.Fixes #25
Test plan
uv run pytest— 291 passed (12 new tests covering records/digest integrity, routing safety invariants, threshold/temperature fitting, replay equivalence, worker protocol)git diff --checkclean; model weights/venv/reports stay under ignoredtraining/.work/Generated with Devin