Skip to content

dataset: use v2 GermanGovService retrieval data - #5323

Merged
Samoed merged 1 commit into
embeddings-benchmark:mainfrom
GioiaZheng:dataset/german-gov-service-v2
Aug 29, 2026
Merged

Samoed merged 1 commit into
embeddings-benchmark:mainfrom
GioiaZheng:dataset/german-gov-service-v2

Conversation

@GioiaZheng

@GioiaZheng GioiaZheng commented Aug 29, 2026 •

Copy link
Copy Markdown
Contributor

Why

GermanGovServiceRetrieval is listed in #3424 as a retrieval task that still uses the old custom loading path. The dataset has already been re-uploaded under the MTEB organization with the standard corpus, queries, and qrels configs.

Refs #3424.

What changed

  • Point GermanGovServiceRetrieval to mteb/GermanGovServiceRetrieval at revision a6cf81304bf8b82d5497a0f7ad9e08399b8f27d3.
  • Remove the custom hash-based load_data() implementation.
  • Let the shared retrieval loader handle the v2 dataset layout.

Validation

  • Rebuilt the old-format IDs from it-at-m/LHM-Dienstleistungen-QA@ed40131b56ce86ce3666f2942953595dd9d29608 and compared them with mteb/GermanGovServiceRetrieval@a6cf81304bf8b82d5497a0f7ad9e08399b8f27d3: corpus, queries, and qrels match exactly (105 corpus docs, 356 queries, 356 qrels).
  • Loaded the v2 dataset through RetrievalDatasetLoader: 105 corpus docs, 356 queries, 356 qrels, no qrels pointing to missing query or corpus IDs.
  • python -m py_compile mteb\tasks\retrieval\deu\german_gov_service_retrieval.py
  • .venv\Scripts\ruff.exe check mteb\tasks\retrieval\deu\german_gov_service_retrieval.py
  • git diff --check

Risk / follow-up

Low. This changes only one German retrieval task to use the verified MTEB-hosted v2 dataset. Follow-up remains for other retrieval tasks in #3424.

@ayush1298

Copy link
Copy Markdown
Collaborator

Can you run evaluation on some model which already has results on these task and compare results, just to check these is not causing any regression in results

@GioiaZheng

Copy link
Copy Markdown
Contributor Author

Thanks, I ran a regression check with an existing results model.

  • Model: sentence-transformers/all-MiniLM-L6-v2 at revision 8b3219a92973c328a8e22fadcfa821b5dc75636a
  • Existing result: embeddings-benchmark/results entry for the old dataset revision ed40131b56ce86ce3666f2942953595dd9d29608 with MTEB 1.12.75
  • This PR: v2 dataset revision a6cf81304bf8b82d5497a0f7ad9e08399b8f27d3 with MTEB 2.20.3
metric existing result this PR
ndcg_at_1 0.37640 0.37640
ndcg_at_3 0.54002 0.54002
ndcg_at_5 0.57882 0.57882
ndcg_at_10 0.61125 0.61125
recall_at_5 0.75281 0.75281
recall_at_10 0.85112 0.85112

So the v2 dataset switch does not appear to introduce a result regression for this task/model.

@ayush1298

Copy link
Copy Markdown
Collaborator

LGTM !
@Samoed Can you check?

@Samoed
Samoed merged commit 211c501 into embeddings-benchmark:main Aug 29, 2026
13 checks passed
@Samoed

Samoed commented Aug 29, 2026

Copy link
Copy Markdown
Member

This dataset uploaded with push_dataset_to_hub, so there shoudn't be problems

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants