Human Grading — Abbreviation Study

Progress & export

Admin view. Grading progress is always visible; the merged export unlocks only after both graders have marked their grading complete.

Progress

GraderGradedStartedLast activityLocked
Grader A346 / 346
7/28/2026, 1:51:11 PM7/28/2026, 6:27:50 PMlocked 7/28/2026, 6:27:53 PM
Grader B346 / 346
7/28/2026, 1:51:44 PM7/28/2026, 6:57:28 PMlocked 7/28/2026, 6:57:52 PM

Human-graded statistics

Each grader's rates across all 346 items, and separately within each prompt variant. Percentages are over items with a value for that field (abbreviation accuracy excludes any items with no abbreviation to score). Grader A and B are shown separately — do not average them; use the merged export for agreement analysis.

Grader A

MetricOverallReal high contextReal low contextSynthetic high context
items graded346116112118
Accuracy — mean
0–2, all items
1.291.931.020.91
Accuracy = 2 (best)
% of scored
49% (n=346)93% (n=116)45% (n=112)9% (n=118)
Accuracy = 0 (worst)
% of scored
20% (n=346)0% (n=116)43% (n=112)19% (n=118)
Clarification — mean
0–2, real_low_context only
1.041.04
Clarification = 2 (explicit ask)
% of scored
38% (n=112)38% (n=112)
Hallucination — mean
0–2, synthetic_high_context only
0.900.90
Hallucination = 2 (false established)
% of scored
21% (n=118)21% (n=118)

Grader B

MetricOverallReal high contextReal low contextSynthetic high context
items graded346116112118
Accuracy — mean
0–2, all items
1.291.961.120.81
Accuracy = 2 (best)
% of scored
50% (n=346)97% (n=116)46% (n=112)6% (n=118)
Accuracy = 0 (worst)
% of scored
20% (n=346)2% (n=116)35% (n=112)25% (n=118)
Clarification — mean
0–2, real_low_context only
1.241.24
Clarification = 2 (explicit ask)
% of scored
57% (n=112)57% (n=112)
Hallucination — mean
0–2, synthetic_high_context only
1.421.42
Hallucination = 2 (false established)
% of scored
47% (n=118)47% (n=118)

Merged export

Both graders are locked. The export merges both score sets by evaluation_id and joins them back to run_id, model_key, variant, and domain using the confidential key.

What the export contains

One row per response — 346 rows carrying all 692 grading decisions, merged on evaluation_id with both graders' original scores side by side. Identity columns: run_id, model_key, parameter_count_b, variant, domain, set_id, prompt_id, abbreviation. Then the three V3 score fields twice, prefixed grader_a_ and grader_b_ (accuracy_score, clarification_score, hallucination_score), plus notes and graded_at. Clarification is only populated for real_low_context items and hallucination only for synthetic_high_context.

Every response is double-graded, so the side-by-side columns support the agreement analysis the V3 spec asks for: raw agreement and Cohen's kappa on Accuracy, Clarification, and Hallucination. Neither original score set is modified — record adjudicated scores separately.