Add judge effects to IRT equation - #2
Open
rohit-ptl wants to merge 1 commit into
Open
Conversation
GIM ability estimates now come from the judge fixed-effect item bank:
y_jk = a_j * (theta - b_j) + gamma_k, sum_k gamma_k = 0
replacing the plain 2PL bank, which silently baked a single judge's
leniency into the item parameters. The five calibrated judges differ in
leniency by up to 1.10 logits, so a bank without a judge term was only
usable with the exact judge it was built on; swapping judges re-ranked
the leaderboard. Publishing one bank plus five centered offsets makes
those judges interchangeable: any single one places a model on the
shared scale, and the scorer subtracts that judge's gamma.
Scoring now requires a calibrated judge. An uncalibrated judge is
rejected when the scorer is constructed, before any inference runs,
rather than producing a biased theta.
irt_params.json v1 -> v2 bank (615 public items) carrying
judge_effects and permitted_judges
irt.py ScoringObservation, score_observations(), gamma
adjustment, item-bank validation; n_items_scored
counts distinct prompts, n_observations the rows
judges.py new registry of the five calibrated judges
metrics.py gim_score() builds judge-labelled observations
scorers.py judge_id / judge_model params, judge metadata
task.py same, defaulting to gemini-3-flash-preview
grader_model is replaced by judge_model; judge_id selects the judge.
Empty completions remain NaN (missing), unchanged.
Verified against the paper pipeline: scoring the 1.09M-row evaluation
set through this scorer reproduces both published ability tables to
floating-point epsilon (max |dtheta| 2.2e-16 over 46 configurations,
zero rank shift) for the 820-item and 615-item banks alike.
Note that the bundled bank is the 615-item public subset, while the
paper reports GIM-820; against the GIM-820 leaderboard the public bank
gives MAE 0.042 and rank shifts up to 5 places.
alexandrerezende
approved these changes
Aug 12, 2026
rohit-ptl
marked this pull request as ready for review
August 12, 2026 20:23
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.
GIM ability estimates now come from a judge fixed-effect item bank:
The previous bank baked one judge's leniency into the item parameters, so it was
only valid with the exact judge it was built on — swapping judges re-ranked the
leaderboard. The five calibrated judges differ in leniency by up to 1.10 logits.
Publishing one bank plus five centered offsets makes them interchangeable: any
single one places a model on the shared ability scale, and the scorer subtracts
that judge's gamma.
Scoring now requires a calibrated judge.
judge_idselects it (defaultgemini-3-flash-preview) andjudge_modelsets the route;grader_modelisgone. An uncalibrated judge is rejected when the scorer is constructed, before
any inference runs, rather than producing a biased theta.
Breaking:
-T grader_model=is removed,GIMScorer.score()requires ajudgeargument, andgim_scoreraises on logs withoutjudge_idmetadata(pre-v2 logs cannot be re-scored). Intentional — a silently biased theta is
worse than a hard failure.