Datasets:
Judge-panel calibration
Last updated: 2026-09-07.
This is the human calibration stage of the LocalGate audit release. It contains 100 answer keys from 90 questions, three human verdicts per answer, and 300 machine verdicts from the three-judge panel. It is separate from the conversion-filter and rewrite-fidelity audits and from the 126,810 bulk judgements.
Load and join
from datasets import load_dataset
repo = "localgate/audit-results"
items = load_dataset(repo, "calibration_items", split="train", token=True)
judgements = load_dataset(repo, "calibration", split="train", token=True)
For a private repository, authenticate with an account that has access. Pin a
Hub commit with revision= when reproducing results. Here train is an archival
storage partition, not a training assignment.
| Configuration | Rows | Unit and key |
|---|---|---|
calibration_items |
100 | Exact input, numbered human labels, and binary majorities; unique key |
calibration |
300 | One judge evaluation per answer; unique (key, judge) |
key is the string question_id:rep, where rep is the Gemma attempt index
0–4. In the
generation dataset
the corresponding field is attempt. All 100 calibration question and response
strings exactly match their corresponding published bulk-generation inputs.
Draw and annotation
The seeded draw (20260815) selected short, long, and list-style edge cases
alongside a category-stratified core: 15 shortest responses, ten longest,
five list-style responses, then five fresh answers per academic category.
The source procedure is
calibration_sample.py.
This deliberately selected set calibrates the grading instrument; it is not an
unbiased sample for estimating overall model accuracy.
Three humans applied the same final-answer/reference rubric as the judges,
with model verdicts concealed during annotation. match maps to 1;
no_match and no_answer map to 0. The human binary 2-of-3 majority is the
comparison target. A three-way human tie still has a defined binary majority;
one such tie occurs in these 100 answers. Human labels occupy stable numbered
columns. Names and annotation notes are omitted; these are not a claim of
irreversible anonymity.
The machine panel used AWS Bedrock models
mistral.ministral-3-14b-instruct, openai.gpt-oss-120b-1:0, and
qwen.qwen3-next-80b-a3b. Settings were temperature 0, a 400-token output cap,
and prompted JSON, with reasoning_effort=low for GPT-OSS. The shared
bulk-grading protocol
contains the exact rubric and output schema; downloadable copies are included
here as system_prompt.txt,
user_template.txt, and
grade.schema.json.
Results and the gate deviation
| Comparison with human binary majority | Scott's π |
|---|---|
| Ministral | 0.7999 |
| GPT-OSS | 0.8599 |
| Qwen | 0.7999 |
| Three-judge binary majority | 0.79992 |
All three individual judges exceeded the 0.60 threshold. The panel agreed with the human majority on 90 of 100 answers, but its exact Scott's π, 1999/2499 = 0.7999199679…, is below the 0.80 panel threshold. Rounding does not make it pass.
Bulk judging proceeded without regrading under the recorded deviation. Eight of the panel's ten disagreements had at least one human agree with the panel. One of the two unanimous-human disagreements involved rejecting $605.92 against a $606 reference, within the rubric's 1% numeric tolerance. These observations contextualize the decision; they do not remove the failed gate or revise the published labels.
The recomputed report records the unrounded panel
statistic, failed gate, per-judge results, confusion counts, and input/source
checks. The scoring implementation is
score_calibration.py.
Data fields
calibration_items
| Field | Meaning |
|---|---|
key, question_id, rep |
Answer key, source question identity, and Gemma attempt index |
question |
Exact question shown for grading |
reference_answer |
Authoritative reference used by the rubric |
model_response |
Exact Gemma response graded by humans and judges |
human_1_verdict, human_2_verdict, human_3_verdict |
Each human's match, no_match, or no_answer verdict |
human_correct |
Binary 2-of-3 human majority, 0 or 1 |
panel_correct |
Binary 2-of-3 machine majority, 0 or 1 |
calibration
| Field | Meaning |
|---|---|
key, question_id, rep |
Answer identity; key is unique only with judge |
judge |
Full bedrock: provider/model specification |
judge_model |
Convenience alias: ministral, gptoss, or qwen |
verdict |
Machine verdict under the same three-way rubric |
verdict_binary |
1 for match; 0 for either other verdict |
reason |
Machine-generated explanation, not verified ground truth |
tokens_in, tokens_out |
Recorded input/output usage for the retained grading operation |
finish_reason |
Gemma generation finish signal; null because original calibration verdicts did not record it |
prompts |
System-prompt-plus-schema fingerprint, e2c597632659 |
template |
User-message-template fingerprint; null in these legacy records |
output_mode |
prompted JSON |
reasoning_effort |
low for GPT-OSS; null means unspecified for the other judges |
Provenance fields were restored from the original judgement records. The
system/schema fingerprint, output mode, and reasoning settings match bulk.
The original calibration records predate the separately recorded user-template
digest (ce0113675bf4). A null digest denotes missing provenance, not proof of
a different template. The system/schema fingerprint does not cover the user
message or provider-added instructions.
Scope and limitations
The 100 answer keys include 75 training, seven validation, and 18 study-test answers under the frozen question split. The 18 test answers were seen during instrument calibration, so they are not an independent held-out calibration evaluation. Exclude calibration question identities, rather than only their individual attempt keys, when claiming independence at the question level.
The target is a three-human majority on a small, deliberately selected set; it is not infallible ground truth. Both human and machine graders can disagree about answer commitment, numeric tolerance, and whether extra detail changes an answer. This calibration does not establish general grader reliability outside the source benchmark and should not be conflated with a later readjudication sample or participant study.