Darsala/GraphRAG-Bench-Georgian

GraphRAG-Bench — Georgian (ქართული)

A complete Georgian translation of GraphRAG-Bench (paper, code), for evaluating Graph Retrieval-Augmented Generation (GraphRAG / KG-RAG) systems and LLMs in Georgian.

Same format as the original: identical file layout, record order, ids, sources, question_types, keys and list lengths, so code written for GraphRAG-Bench runs unchanged. Every text field is in Georgian: the corpora (context), question, answer, evidence, evidence_relations and evidence_triple.

File	Records
Datasets/Corpus/novel.json	20 novels
Datasets/Corpus/medical.json	1 medical corpus
Datasets/Questions/novel_questions.json	2,010 questions
Datasets/Questions/medical_questions.json	2,062 questions

⚠️ Two quality tiers: read this before reporting results

Translation was done per corpus, so each novel together with its questions comes from a single source:

• High quality (Claude tier): 14 of the 20 novels, 1,314 questions. Translated by Claude Fable 5.1 or Claude Opus 5.5. These models scored 87–90/100 in blind multi-judge MQM evaluation on a held-out sample, with almost no major errors.
• Lower quality (kona tier): the other 6 novels and the whole Medical subset, 696 + 2,062 questions. Translated by the open-weight kona2-12B, which scored 43/100 on the same test (about 7 major errors per passage: wrong word senses, garbled names, occasional omissions, case errors). Structural and automatic checks pass, but expect noise. Use these corpora for coverage, not as gold data.

For clean model comparisons, use the high-quality novel subset:

import json
prov = json.load(open("translation_provenance.json"))["corpora"]
high = {c for c, v in prov.items() if v["tier"] == "claude"}
qs = [q for q in json.load(open("Datasets/Questions/novel_questions.json")) if q["source"] in high]
docs = [d for d in json.load(open("Datasets/Corpus/novel.json")) if d["corpus_name"] in high]

High-quality novels: Novel-30752, Novel-51410, Novel-58553, Novel-10762, Novel-26183, Novel-29973, Novel-47676, Novel-10321, Novel-47558, Novel-25646, Novel-54537, Novel-40700, Novel-4128, Novel-44557.

Corpus	Quality	Translated by	Questions
Medical	⚠️ lower	kona2-12B	2062
Novel-30752	✅ high	Claude Fable 5.1	87
Novel-51410	✅ high	Claude Fable 5.1	78
Novel-58553	✅ high	Claude Fable 5.1	112
Novel-10762	✅ high	Claude Fable 5.1	78
Novel-26183	✅ high	Claude Fable 5.1	96
Novel-29973	✅ high	Claude Fable 5.1	102
Novel-10146	⚠️ lower	kona2-12B	126
Novel-47676	✅ high	Claude Fable 5.1	81
Novel-10321	✅ high	Claude Fable 5.1	71
Novel-8559	⚠️ lower	kona2-12B	137
Novel-47558	✅ high	Claude Fable 5.1	108
Novel-25646	✅ high	Claude Opus 5.5	104
Novel-5956	⚠️ lower	kona2-12B	86
Novel-2544	⚠️ lower	kona2-12B	114
Novel-41603	⚠️ lower	kona2-12B	103
Novel-54537	✅ high	Claude Opus 5.5	120
Novel-10356	⚠️ lower	kona2-12B	130
Novel-40700	✅ high	Claude Opus 5.5	102
Novel-4128	✅ high	Claude Opus 5.5	72
Novel-44557	✅ high	Claude Opus 5.5	103

The per-corpus details are in translation_provenance.json.

How it was translated

• Chunking: corpora were split at sentence boundaries into chunks of about 6,000 characters, and each chunk was translated with about 700 characters of preceding context. Questions were translated in groups of 10, using structured JSON output.
• Per-corpus glossaries (glossaries/): built from the full corpus plus all of its questions, answers and evidence. They fix one Georgian form for every named entity and domain term, so that entities match across chunks and questions, which knowledge-graph construction needs.
• Prompting (Claude tier): a Georgian-specific checklist covering ergative/dative case agreement with verb series, Georgian clock idioms, preserving markers such as <DW53>/[12], and medical terminology. This prompt was selected by held-out A/B evaluation.
• Validation: every unit was checked automatically for length ratio, untranslated English, mixed-script words, leaked tags, lost markers, triple format (head,relation,tail) and list lengths. Failures were re-translated. The kona tier also has stricter length and repetition guards and several fallbacks: smaller segments, masked markers, and triples requested as structured parts.
• Masking tokens: markers such as <DW52> are inherited from the original dataset and are kept unchanged.

Evaluation code

Evaluation/ and Examples/ are copied from the original. One change was needed for Georgian: Evaluation/metrics/rouge.py uses a Unicode tokenizer. The original rouge_score tokenizer keeps only [a-z0-9], which silently scores every Georgian answer as 0. The LLM-judged metrics work as they are. Use an LLM judge that is strong in Georgian.

License and citation

MIT, the same as the original GraphRAG-Bench. Please cite the original benchmark:

@article{xiang2025use,
  title={When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation},
  author={Xiang, Zhishang and Wu, Chuanjie and Zhang, Qinggang and Chen, Shengyuan and Hong, Zijin and Huang, Xiao and Su, Jinsong},
  journal={arXiv preprint arXiv:2506.05690},
  year={2025}
}