One config, three splits, one schema. Everything needed to build a prompt and a training target. No question text: the template is chosen at inference time.

split	rows	traced	what it is
`train`	272,688	272,688	training corpus (1,321 untraced rows dropped)
`test`	6,112	6,007	the benchmark — score here
`test_old`	3,726	3,487	previous benchmark, usable subset only

Test splits keep untraced rows; the model is evaluated by generating.

Leakage — checked on protein id and sequence md5: train ∩ test = 0/0 · train ∩ test_old = 0/0 · test ∩ test_old = 968 ids (see §4).

1. 37,947 train rows have `final_answer = NULL`. They are not broken.

The grounding split (is_grounding=True, prompt_version="v6"): the opening paragraph of a trace, ~2,500 chars of architectural narration, stopping before any conclusion or GO term. It teaches the model how to open a trace.

Masking matters. Trained naively it teaches the model to stop after one paragraph, one level below where it looks. The supervised span must end on the paragraph's last token and include neither </think> nor <|im_end|>. Filter on is_grounding to exclude them.

2. Use `go_pred_ep{25,50,75}`. Never a bare `go_pred`.

Rendered here from the raw go_{mf,bp,cc}_gogpt_ep* columns, which also ship. The upstream go_pred string column is empty across the extra corpus and carries a different, older predictor elsewhere — the field most likely to be silently wrong if trusted.

Pick one checkpoint and stay on it: a model trained against one predictor must be evaluated against the same one.

3. Four feature blocks. Not interchangeable.

column	contents
`features_formatted`	everything: InterProScan + UniProt curation + PDB
`features_formatted_iscan`	`src ∈ {interpro_match, interpro_site}` only
`interpro_formatted`	IPR-bearing subset of the full block
`interpro_formatted_iscan`	IPR-bearing subset of the InterProScan block

_iscan is what a novel sequence can be given — 78% of all feature records — and what deployment-condition models train on. The full block adds BINDING, ACT_SITE, MUTAGEN, DISULFID, TRANSMEM, MOD_RES: facts that exist only because a curator read a paper. Training on the full block and serving on _iscan is train/serve skew.

Raw features structs ship too, so any other rendering is reproducible.

Spans are per-MATCH, not InterPro's union span. An entry matched several times carries each: IPR000000 name [11-92] [140-259]. The union merges tandem repeats into a phantom domain — measured here, 27.3% of multi-match entries have a union covering residues no match reports (median 33% of the span; worst case a 70-residue repeat rendered as 27,540).

4. `test_old` is filtered, and overlaps `test`.

The previous benchmark was 8,084 rows. 4,358 are excluded upstream — uniprot_deleted 2,475, tier1_retracted_since_gaf226 854, immature_annotation 382, no_tier1_any_release 298, train_subsequence 244 — so scoring them means scoring against labels UniProt withdrew. The 3,726 kept are exactly excluded_reason IS NULL, which agrees with "non-empty scored_aspects" on all 8,084 rows.

968 of them are also in test (all test_quality_basis == "benchmark", promoted into the current 6k). The two test splits are not independent.

test_quality is a separate axis and grades a row that IS scoreable: the usable 3,726 are 2,479 high / 1,247 low.

Provenance

source on every row: v2 (132,651 train) · v2_extra (102,090) · v2_extra_ground (37,947).

Traces written by moonshotai/kimi-k3 at high reasoning effort, then a three-tier QC pass — GO id/name repair against go-basic.obo, section refiling, restoration of supplied InterPro entries and partners — audited against each record's raw generation: zero invented GO ids, InterPro accessions or partners; zero stray control delimiters.

Labels, scored_aspects and the release strings (uniprot_release, interpro_release, gaf_release, obo_version) come from wanglab/bioreason-pro-v2, carried unmodified.

One known gap. features_formatted is rendered with the current renderer, which fixes an earlier bug that duplicated a position in scattered-site bullets ([residue: HHEE] where three residues are meant). The v2 traces were written before that fix, so 6.84% of source=v2 rows differ from the block their own trace was generated against; v2_extra is 0.01% and v2_extra_ground 0.00%. The shipped column is the corrected one. This matters only if you are reproducing a trace's input prompt byte-for-byte.