anonymousresearch123/deka-retrieval

Deka retrieval

A legal case retrieval (LCR) benchmark on Thai Supreme Court decisions (คำพิพากษาฎีกา, deka): given a case, retrieve the prior cases that bear on it out of the full 9,998-case pool. Thailand is a civil-law jurisdiction, so relevance is defined over case components — charge, facts, issue, outcome — rather than citation links between judgments, and that criterion is what yields the four label quadrants below. Candidates were mined with a dense retriever and adjudicated by an LLM judge, and every anchor carries three query registers. Anonymous release accompanying a paper under review.

Files

file	rows	schema
corpus.jsonl	9,998	_id, deka_code, long_text, metadata{deka_no, year, fact, reasoning, laws, crimes, case_type, case_type_group, issue, decision}
queries.jsonl	1,381	_id, long_text, lay_query, lawyer_query, metadata{…, split}
labels.jsonl	1,381	query_id, deka, split, mined, easy_pos[], hard_pos[], hard_neg[], easy_neg[], anomaly[]

Every entry in a label list is {candidate_id, deka, score, why, mining_query} — candidate_id joins to corpus._id, mining_query records which query text surfaced that candidate.

Query registers

Each anchor carries three usable query forms, which retrieve very differently:

• lay_query — how a layperson would describe the situation (short, non-technical)
• lawyer_query — how a lawyer would phrase it (statutory language)
• metadata.fact — the case's own fact section, i.e. the symmetric case-to-case setting

Splits

labels.jsonl is authoritative. Anchors are split train 1,176 / eval 100 / test 105.

Excluded: 191 anchors marked keep_test are not part of this release. In queries.jsonl, 100 rows whose metadata.split disagreed with labels.jsonl have been corrected to the label value, so the two files now agree row for row.

The 9,998-case corpus is untouched and shared by every split.

Heuristic case criteria

• easy_pos — same charge, same decision outcome
• hard_pos — same charge, different outcome
• hard_neg — retrieved as similar, judged not relevant
• easy_neg — clearly unrelated
• anomaly — candidates the judge could not place cleanly; not relevant by default

Roughly one candidate per list per anchor (1,401 entries per list across 1,381 anchors); 175 anchors have no anomaly entry.

Caveats

• Positives are mined, not exhaustively annotated. A case absent from easy_pos/hard_pos is not certified irrelevant, so recall computed on this data is a lower bound.
• Judgements come from an LLM judge over retrieved candidates; they inherit both the retriever's blind spots and the judge's.
• issue[] and decision were extracted from long_text by rule, not by hand.
• Decisions are published court documents, but they name parties and describe offences — treat the text as sensitive and do not use it to profile individuals.
• Absolute retrieval quality is low (best dense R@10 ≈ 31 on fact queries); this is a research bed, not a solved task.

Loading

from datasets import load_dataset

corpus  = load_dataset("anonymousresearch123/deka-retrieval", "corpus",  split="train")
queries = load_dataset("anonymousresearch123/deka-retrieval", "queries", split="train")
labels  = load_dataset("anonymousresearch123/deka-retrieval", "labels",  split="train")

Each config is a single file; filter metadata.split / split for train / eval / test.