@context dict | @type string | @id string | name string | description string | conformsTo string | license string | url string | version string | datePublished string | creator dict | citation string | distribution list | rai:dataBiases string | rai:dataUseCases string | rai:personalSensitiveInformation string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{
"@language": "en",
"@vocab": "https://schema.org/",
"citeAs": "cr:citeAs",
"column": "cr:column",
"conformsTo": "dct:conformsTo",
"cr": "http://mlcommons.org/croissant/",
"rai": "http://mlcommons.org/croissant/RAI/",
"data": {
"@id": "cr:data",
"@type": "@json"
},
"dataType": {
"@id": ... | sc:Dataset | https://example.com/datasets/paper5-contamination-db | paper5-contamination-db | Per-example LLM contamination scores for 14-18 open-weight models across 13 benchmarks under two calibration regimes (v1: cross-benchmark negatives [deployment-invalid, replication-only]; v2: same-benchmark negatives [deployment-valid]). | http://mlcommons.org/croissant/1.0 | https://creativecommons.org/licenses/by/4.0/ | https://example.com/datasets/paper5-contamination-db | 1.0-neurips2026 | 2026-04-23T09:55:28.467500+00:00 | {
"@type": "Organization",
"name": "anonymous-for-review"
} | Anonymous, 'The Calibration Mismatch in LLM Contamination Detection', NeurIPS 2026 D&B submission. | [
{
"@type": "cr:FileObject",
"@id": "ensemble_v2",
"contentUrl": "ensemble_v2.pkl",
"sha256": "TBD",
"encodingFormat": "application/python-pickle",
"description": "v2 (same-benchmark negatives) calibrated ensemble — safe for deployment.",
"includes": null
},
{
"@type": "cr:FileObj... | v1 scores exhibit a systematic domain-benchmark confound; see paper §4-5. v2 scores are the recommended deployment target. Scores are uncalibrated against closed-API LLMs (GPT-4/Claude/Gemini); methodology is demonstrated on open-weight models only. | Intended for: (a) replication of the v1/v2 calibration-mismatch finding; (b) benchmark-audit workflows that feed per-example scores into within-difficulty-stratified accuracy-gap tests (§5.2 of the paper). Not intended for: hard thresholding of individual items without the accompanying accuracy-gap validation. | No PII. All benchmark items are from public evaluation datasets with open licenses. |
No dataset card yet
- Downloads last month
- 2