Datasets:
Download metadata/source-format-sensitivity-specification.json from superlinked/train-a-router-bbeh: direct link, hf CLI and curl.
- Browser
- Download file 3.02 kB
-
https://huggingface.co/datasets/superlinked/train-a-router-bbeh/resolve/main/metadata/source-format-sensitivity-specification.json
- Command line
-
hf download hf://datasets/superlinked/train-a-router-bbeh/metadata/source-format-sensitivity-specification.json
-
curl -L -o source-format-sensitivity-specification.json https://huggingface.co/datasets/superlinked/train-a-router-bbeh/resolve/main/metadata/source-format-sensitivity-specification.json
3.02 kB
| { | |
| "basis": "Public question text only; no acquired response, gold label, model identity or correctness used to identify ambiguity.", | |
| "final_sensitivity": { | |
| "excluded_ids": [ | |
| "bbeh:disambiguation_qa:051", | |
| "bbeh:disambiguation_qa:052" | |
| ], | |
| "n_primary": 975, | |
| "n_sensitivity": 973, | |
| "name": "exclude_source_ambiguous_options", | |
| "rule": "Evaluate the same frozen policies and per-row costs on final rows except these fixed source-ambiguous IDs. Recompute the same aggregate metrics and uncertainty protocol; report both denominators. No refitting, recalibration, threshold selection or answer regeneration." | |
| }, | |
| "fixed_before_final_outcomes": true, | |
| "format": "bbeh-source-format-sensitivity-v1", | |
| "limitations": "A duplicated label can encode two different meanings and identical text can occupy two labels. Upstream label-only references cannot resolve these source defects. Primary scores retain that limitation; the separate sensitivity exposes its final-evaluation effect.", | |
| "normalizer_sha256": "258a60874b11b5bbeefe2a4fe304dfcc2da341eba995258262ce1e5931ef5772", | |
| "primary_population_and_reference_bytes_unchanged": true, | |
| "primary_role_counts": { | |
| "calibration": 385, | |
| "development": 589, | |
| "final": 975, | |
| "train": 1969 | |
| }, | |
| "source_ambiguous_rows": [ | |
| { | |
| "duplicate_labels": [], | |
| "duplicate_text_labels": [ | |
| [ | |
| "B", | |
| "D" | |
| ] | |
| ], | |
| "family": "2d6679ee5418599e7e79d8eada436d3251c513fd874df67786f41f63e014ab88", | |
| "id": "bbeh:disambiguation_qa:018", | |
| "index": 18, | |
| "request_sha256": "d43a5ad79fd64b28f935dfdd5e308d504200060111f51e7b0bbdfd730927f5be", | |
| "retained": true, | |
| "role": "train", | |
| "task": "disambiguation_qa" | |
| }, | |
| { | |
| "duplicate_labels": [ | |
| "E" | |
| ], | |
| "duplicate_text_labels": [], | |
| "family": "70b9e67438ab0f8bd2e92b6196f9fdb4b82d2011167faddd7115fa9c09132fd8", | |
| "id": "bbeh:disambiguation_qa:051", | |
| "index": 51, | |
| "request_sha256": "3434899c00f908b6f27290e90df888a21c23b739e79357b4d4cad0ecd4046324", | |
| "retained": true, | |
| "role": "final", | |
| "task": "disambiguation_qa" | |
| }, | |
| { | |
| "duplicate_labels": [], | |
| "duplicate_text_labels": [ | |
| [ | |
| "C", | |
| "D" | |
| ] | |
| ], | |
| "family": "fd4491ee83e2c6a26a27f2d38f4056681bbf0564742eae2e3a6e8f13694948e2", | |
| "id": "bbeh:disambiguation_qa:052", | |
| "index": 52, | |
| "request_sha256": "f443e7d31db4fabcd10704f8bd58bdf354a5b847c6803fcad013b5f16bfd7b47", | |
| "retained": true, | |
| "role": "final", | |
| "task": "disambiguation_qa" | |
| } | |
| ], | |
| "source_audit_sha256": "bf618b8be32a6391094b984919f802c8f0dd5c5c98417e8475764c56ee579cdd", | |
| "source_plan_sha256": "b15e9c2d9db284fe170e0460550f02fc2cb48f2b9234b697d181716264984a53", | |
| "training_ambiguity_policy": "Retain disambiguation_qa:018 in the fixed training population with the original source label semantics; do not repair its label or choose an option from model outputs." | |
| } | |