Anirbanbhk/german-public-sector-c_dbr

Dataset Card for publicsectorcdbrQA

Dataset Summary

publicsectorcdbrQA is a German-language QA dataset for public-sector and legal-administrative content. The dataset includes question, answer, source context, and LLM-as-a-judge quality metadata.

Dataset Files

• —Main file: publicsectorcdbrQA.jsonl
• —Base evaluated file: cdbrevaluatedtop20_percent.jsonl
• —Suggested split mapping: train only

Record Counts

• —publicsectorcdbrQA.jsonl: 1981 rows
• —cdbrevaluatedtop20_percent.jsonl: 1981 rows

The row counts match, indicating publicsectorcdbrQA.jsonl is based directly on the top-20-percent evaluated c_dbr subset.

Source Corpus

• —Source collection: c_dbr public-sector/legal-administrative documents used in this repository pipeline.
• —Immediate parent dataset file: cdbrevaluatedtop20_percent.jsonl

How QA Pairs Were Created

1. 1.Source ingestion: Documents from the c_dbr corpus were processed into text chunks for QA generation.
2. 2.Synthetic QA generation: Candidate German question-answer pairs were generated from those chunks.
3. 3.LLM-as-a-judge evaluation: Each QA pair was scored using three metrics:
4. 4.Completeness of Answer
5. 5.Clarity and Conciseness
6. 6.Context Specificity
7. 7.Aggregation: A per-record aggregated_score was computed from valid metric values.
8. 8.Top-k filtering: The top 20 percent results were retained in cdbrevaluatedtop20_percent.jsonl.
9. 9.Export: The filtered set was saved as publicsectorcdbrQA.jsonl.

Data Fields

Each JSONL record contains the following fields:

• —testcaseid: Integer identifier from the evaluated source.
• —question: German question.
• —answer: German answer.
• —context: Source passage used during generation/evaluation.
• —Completeness of Answer_score: Float score in [0, 1].
• —Completeness of Answer_reason: Judge explanation text.
• —Clarity and Conciseness_score: Float score in [0, 1].
• —Clarity and Conciseness_reason: Judge explanation text.
• —Context Specificity_score: Float score in [0, 1].
• —Context Specificity_reason: Judge explanation text.
• —aggregated_score: Mean of valid metric scores used for ranking/filtering.
• —validmetricscount: Number of valid metric values.

Evaluation Tasks This Dataset Supports

• —German QA evaluation on legal/public-administration passages.
• —Retrieval-augmented QA benchmarking where context grounding is required.
• —Ranking and filtering experiments using aggregated_score and per-metric diagnostics.
• —LLM answer-quality evaluation studies using judge reasons and metric scores.

Intended Use

• —Fine-tuning or evaluation of German QA models in public-sector/legal domains.
• —Validation of context-grounded answer generation.

Out-of-Scope Use

• —Legal advice or legally binding interpretation.
• —Fully automated high-stakes decisions without human review.

Quality Notes and Limitations

• —Metric scores are model-based and may contain bias or inconsistency.
• —Some answers can still include minor context drift despite top-20 filtering.
• —The dataset reflects c_dbr domain characteristics and may not generalize broadly.

Licensing

License is currently marked as other.

Before publishing to Hugging Face, replace with the exact license identifier that applies to both source materials and derived annotations.

Citation

If you publish this dataset, add a formal citation entry here.

Example placeholder:

@misc{publicsectorcdbrqa2026, title = {Public Sector cdbr QA (German)}, author = {Your Name or Team}, year = {2026}, howpublished = {Hugging Face Dataset Repository} }