--- license: cc-by-4.0 language: - en configs: - config_name: casino data_files: - split: test path: data/casino.jsonl - config_name: contractnli data_files: - split: test path: data/contractnli.jsonl --- # MediationBench external-source pilot Frozen sample: `external-200-v1`. Included cases: **200**. Source counts: `{"casino": 100, "contractnli": 100}`. These are deliberately varied source-labeled cases, not a representative sample of product traffic. Only CaSiNo consists of conversations: ContractNLI uses contracts and hypotheses. Source labels are not newly human-adjudicated HAI readiness, consent, authorization, or completion judgments. Heuristic difficulty is a frozen selection proxy using source features and annotations, not measured judge difficulty. ## Files and use Select a dataset configuration to load its cases. Each row contains source provenance, heuristic difficulty, a judge-input envelope, and source reference labels. Feed only `judge_input` to models. Reference answers, source IDs, difficulty and sampling rationale must stay outside model inputs. The Hub `test` split names this pilot's evaluation subset; original source splits remain in each record and are not newly held-out data. `sample-manifest.json` records the frozen sample and original file checksums; `bundle-manifest.json` records checksums for the exact published files. Judge outputs and private simulator material are not part of this dataset. ## Sampling and interpretation Selection is frozen before evaluation; the sample manifest records its seed, source revisions, source hashes, strata, counts and selection policy. Do not replace cases after observing judge performance. Report each source task and its coverage separately; their labels do not define one pooled mediation score. Public-source pretraining contamination remains possible. ## Licenses, citations and modifications [LICENSES.md](LICENSES.md) supplies source-specific licenses, bibliographic attribution and modification notices. Each source retains its own terms.