docs/schema.md

Schema

The release has four configs. Ids are content hashes, so anyone holding the text can recompute them; see corpus, queries and test split.

config	split	rows
corpus	train	3,525,546
queries	train	1,853,488
queries	test	6,173
judgments	train	43,508,511
qrels	test	74,519

The corpus serves both splits. Hugging Face names the only split of a config train.

corpus

column	type	meaning
did	string	"d_" + sha256(lowercase(normalized_text))[:24]
text	string	normalized document text, with the source page's own question removed
n_chars	int32	length of text

queries

column	type	meaning
qid	string	training queries: "p_" + sha256(raw_page_text)[:24], the id of the source page the query was extracted from (raw upstream content, no normalization). Test queries: "q_" + sha256(lowercase(normalized_text))[:24], a hash of the query text normalized as for document ids
text	string	the query
source_did	string	the document the query was extracted from, with its own question removed. When near-duplicate removal took that page out, the surviving copy, if the LLM judge found it a full answer (208,303 queries). Null for the other 64,546 training queries (3.5%) and for all test queries
from_page_question	bool	true if the query is the page's own question, minimally edited; false if the model composed it. Null for test queries

judgments

One row per judged (training query, document) pair.

column	type	meaning
qid, did	string	the pair
dense_rank	int8	rank in the dense route's top 100; null if the document is not in it
bm25_rank, ngram5_rank	int8	rank in the BM25 or word 5-gram route's top 100; null if the document is not in it
fused_rank	int8	position in the fused top 20; null if the document is not in it
rrf_score	float64	the fusion score 2 / (60 + dense_rank) + 1 / (60 + bm25_rank) + 1 / (60 + ngram5_rank), where a missing route adds 0; null if no route retrieved the document
stage	string	the group of candidates the pair belongs to; see below
label	int8	0 / 1 / 2; null when the query was judged unevaluable
query_evaluable	bool	false when the LLM judge found the query itself uninterpretable
is_source_page	bool	the document is the query's source_did
judge	string	llm or rm8b, whichever produced label
rm8b_score	float32	the 8B model's raw score, on every pair it scored, including pairs whose label comes from the LLM
dense_score	float32	cosine similarity from the dense retriever, for documents in its top 100
reranker_score	float32	zero-shot score from Qwen3-Reranker-4B (logit of yes minus no), for dense ranks 6–20; null elsewhere
answer_validity, reason_codes, confidence, rationale		the LLM judge's structured output; null on 8B rows
rubric_sha256	string	SHA-256 of the judge prompt, where it was recorded: on LLM-labeled candidates and on the LLM labels of surviving copies; null on 8B rows and on the other source_page rows

stage	pairs	judge
dense_top5	dense ranks 1–5	LLM
rerank_top5	dense ranks 6–20, the reranker's top 5	LLM, also scored by the 8B
long_doc	dense ranks 6–20, longer than 8,144 tokens	LLM
short_doc	dense ranks 6–20, all other documents	8B
fused_long_top5	fused-only, longer than 8,144 tokens, fused positions 1–5	LLM
fused_long	fused-only, longer than 8,144 tokens, fused positions 6–20	8B
fused_short	fused-only, up to 8,144 tokens	8B
source_page	the query's own source page (or the surviving copy that stands in for it), labeled outside retrieval, when no stage above covers it	LLM

Fused-only means in the fused top 20 but not in the dense top 20; see judging.

Rules.

• When both judges labeled a pair, label and judge come from the LLM, and rm8b_score is kept.
• An 8B label applies the default cuts 0.5 and 1.5 to rm8b_score. Use the raw score to pick your own threshold.
• No training qid overlaps the test queries: not verbatim, not at word 5-gram Jaccard ≥ 0.50, and not as a semantic duplicate judged to be the same problem.

qrels

The relevant documents of each test query.

column	type	meaning
qid	string	the test query
did	string	a relevant document
label	int8	always 2: both judges, or the arbiter, found that the document answers the same problem correctly