CoolFace
Datasetpublic

temsa/finephrase-gov-reranker-hard-k10-v1

FinePhrase Gov Reranker Hard K10 v1 Hard reranker dataset derived from HuggingFaceFW/finephrase for Irish-government-like search. Each row is a grouped reranker example: one query one relevant passage from a public-sector page nine hard negatives from the same filtered gov-like domain pool The candidate set is fixed at 10, matching the current serving target of reranking only the top 10 candidates after first-stage retrieval / RRF. Source filtering Rows are… See the full description on the dataset page: https://huggingface.co/datasets/temsa/finephrase-gov-reranker-hard-k10-v1.

sourceHugging Faceodc-byupdated 6mo agoView on Hugging Face
0likes16downloads
Dataset Card

FinePhrase Gov Reranker Hard K10 v1

Hard reranker dataset derived from HuggingFaceFW/finephrase for Irish-government-like search.

Each row is a grouped reranker example:

  • —one query
  • —one relevant passage from a public-sector page
  • —nine hard negatives from the same filtered gov-like domain pool

The candidate set is fixed at 10, matching the current serving target of reranking only the top 10 candidates after first-stage retrieval / RRF.

Source filtering

Rows are filtered to public-service and Irish government related hosts such as:

  • —gov.ie
  • —citizensinformation.ie
  • —hse.ie
  • —oireachtas.ie
  • —related historical/government subdomains captured in the source snapshot

Construction

  • —source dataset: HuggingFaceFW/finephrase
  • —positives: extracted FAQ-like public-domain question/passage pairs
  • —negatives: lexical shortlist followed by teacher-selected hard negatives
  • —teacher used for hard-negative selection: temsa/search-reranker-broad-v1 q8
  • —candidate count: 10

Split sizes

  • —train: 123
  • —validation: 37
  • —test: 27

Intended use

  • —K=10 reranker evaluation for Irish gov search
  • —hard-negative finetuning for smaller CPU rerankers
  • —overfit checking against FAQ-like public-service queries

Attribution

This dataset is derived from HuggingFaceFW/finephrase and keeps its odc-by attribution requirement. Preserve attribution to the original dataset authors and this derived packaging.