---
license: cc-by-4.0
language:
- de
task_categories:
- text-retrieval
tags:
- mteb
- german
- retrieval
- reranking
- benchmark
size_categories:
- 10K<n<100K
pretty_name: MTEB-DE — German Embedding/Retrieval Benchmark
configs:
- config_name: germandpr-corpus
  data_files:
  - split: corpus
    path: germandpr/corpus/corpus.parquet
- config_name: germandpr-queries
  data_files:
  - split: train
    path: germandpr/queries/train.parquet
  - split: dev
    path: germandpr/queries/dev.parquet
  - split: test
    path: germandpr/queries/test.parquet
- config_name: germandpr-qrels
  data_files:
  - split: train
    path: germandpr/qrels/train.parquet
  - split: dev
    path: germandpr/qrels/dev.parquet
  - split: test
    path: germandpr/qrels/test.parquet
- config_name: jobs-corpus
  data_files:
  - split: corpus
    path: jobs/corpus/corpus.parquet
- config_name: jobs-queries
  data_files:
  - split: train
    path: jobs/queries/train.parquet
  - split: dev
    path: jobs/queries/dev.parquet
  - split: test
    path: jobs/queries/test.parquet
- config_name: jobs-qrels
  data_files:
  - split: train
    path: jobs/qrels/train.parquet
  - split: dev
    path: jobs/qrels/dev.parquet
  - split: test
    path: jobs/qrels/test.parquet
- config_name: jobs-hard_negatives
  data_files:
  - split: train
    path: jobs/hard_negatives/train.parquet
  - split: dev
    path: jobs/hard_negatives/dev.parquet
  - split: test
    path: jobs/hard_negatives/test.parquet
---

# MTEB-DE — German Embedding/Retrieval Benchmark

A consolidated, reproducible German retrieval benchmark. v0 ships the
**retrieval** task type across 2 configs; reranking / STS / clustering
follow. Loaded via, e.g.,
`load_dataset("mischeiwiller/mteb-de", "germandpr-queries")` (BEIR/MTEB layout: each task =
`<config>-corpus` / `-queries` / `-qrels` [ + `-hard_negatives`]).

## Dataset Summary

- **`germandpr`** — German open-domain QA passage retrieval (deepset GermanDPR via the CC-BY-4.0 MTEB mirror), re-split train/dev/test by query id (seed 42); the corpus is shared across splits per BEIR/MTEB index semantics.
- **`jobs`** — Net-new hard German retrieval: a job title must retrieve the canonical occupation/skills profile of the same posting. Derived from Project-1 official-API job data (no verbatim postings), PII-scrubbed, with mined hard negatives.

## Configurations

Each retrieval task is published as separate corpus / queries / qrels configs
(plus `hard_negatives` where mined), so the index is shared across the
train/dev/test query splits — standard BEIR/MTEB retrieval semantics.

## Provenance & Licensing

All configs are CC-BY-4.0. We publish **derived** fields only — no verbatim,
ToS-restricted source text. Source dataset and pinned revision per config:

| Config | Upstream source | Loaded from | Revision | License |
|---|---|---|---|---|
| `germandpr` | `deepset/germandpr` | `mteb/GermanDPR` | `64a4860e55` | cc-by-4.0 |
| `jobs` | `mischeiwiller/german-job-postings` | `mischeiwiller/german-job-postings` | `d49f1e4243` | cc-by-4.0 |

## Splits

Queries (and their qrels / hard negatives) are partitioned **by query id** into
`train` / `dev` / `test` at 0.8 / 0.1 / 0.1 with a fixed seed (42). The corpus is
a single shared `corpus` split.

## Metrics

- `germandpr`: nDCG@10 / MRR@10
- `jobs`: nDCG@10 / MRR@10

## Hard-Negative Mining

The `jobs` config ships mined hard negatives. Procedure: encode queries and
passages with `intfloat/multilingual-e5-base`@`d128750597` (e5 `query:` / `passage:`
prefixes), rank by cosine similarity, take top-10, drop the
labelled positive(s), drop candidates within 0.05 of
the positive score (false-negative guard), and keep up to
5. Full write-up + spot-audit: see the project
`notes/hard-negative-mining.md` and `notes/hard-negative-audit.md`.

## Limitations

- `germandpr` is from 2021 and small (~1k queries); included for continuity.
- `jobs` passages are composed from structured ESCO fields (occupation + skills,
  region, seniority, KldB) because the upstream `description_derived` is
  currently unpopulated; ESCO labels are English, making `jobs` a cross-lingual
  German-title → English-profile task in v0.
- Hard negatives are mined automatically and spot-audited, not exhaustively
  human-verified; residual false negatives are possible.

## Citation

Derived from [`deepset/germandpr`](https://huggingface.co/datasets/deepset/germandpr)
(via the MTEB mirror) and [`mischeiwiller/german-job-postings`](https://huggingface.co/datasets/mischeiwiller/german-job-postings).
A canonical Croissant descriptor is auto-served by the Hub at the dataset's
`/croissant` endpoint; a generated copy ships as `croissant.jsonld`.