open-athena/MarinFold/data/document_structures/contacts_and_coordinates_v1 - Storage Bucket

contacts-and-coordinates-v1 protein-document dataset

Training documents in the contacts-and-coordinates-v1 format (single-chain protein → sequence section + side-chain contact statements + 3D coordinate mention events at 1 Å resolution), generated from timodonnell/afdb-24M by calling MarinFold's contacts_and_coordinates_v1 generator (no re-implementation). See the format spec: marinfold/.../document_structures/contacts_and_coordinates_v1.

Produced by experiment exp105 (marinfold @ 02e9202), reusing the exact same proteins / rounds / splits as contacts-v1 — exp53's selection manifest was reused verbatim, so there is no new selection and the two corpora are aligned row-for-row on (entry_id, round, split).

Format (1 Å coordinates, budget-filling)

context_length = 32768, max_depth = 3 (1 Å resolution). Unlike contacts-v1 (8192 tokens, mostly sequence + contacts), coordinate mention events are sampled with replacement until the token budget fills, so every document reaches ~32,766 tokens regardless of protein length (measured min 32,764, max 32,768). This makes the corpus ~30× larger in tokens than contacts-v1 for the same proteins.

Structure (rounds)

Up to 5 examples per structural cluster, organised into pLDDT rounds: round = 0 is the highest-pLDDT representative of each struct_cluster_id, round = 1 the next, etc. Structural clusters with fewer than 3 usable members are dropped (so round-0 == round-1 == round-2 in size). Shards are laid out round-descending (the shard index encodes order: round-4 first … round-0 last) so a sequential reader trains on the highest-quality data last. The round column is authoritative regardless of file order.

split (train/val/test) is inherited from afdb-24M and is cluster-consistent (a whole structural cluster stays in one split → no train/test leakage).

Layout

<split>/ccoord_v1-{shard:05d}-of-{total:05d}.parquet   # ≤2000 rows/shard, round-descending index
tokenizer/                                             # contacts-and-coordinates-v1 tokenizer (3847 tokens)

Counts

Same proteins as contacts-v1 (exp53). 0 generation drops — every selected (entry, round) serialized (per-split records_in == records_out, and per-round counts match the exp53 manifest exactly).

split	round 0	round 1	round 2	round 3	round 4	total	tokens
train	941,028	941,028	941,028	719,519	587,079	4,129,682	135.3 B
val	9,558	9,558	9,558	7,316	5,964	41,954	1.375 B
test	9,468	9,468	9,468	7,248	5,915	41,567	1.362 B
total	960,054	960,054	960,054	734,083	598,958	4,213,203	~138.0 B

Mean num_tokens ~32,766/doc (min 32,764, max 32,768) — every document fills the 32,768 budget. train is 303 GiB of parquet, written round-descending by shard index. All three splits byte-verified: sha1 == sha1(document) and per-round counts exact.

Columns

document (token string) · structure ("contacts-and-coordinates-v1") · entry_id · seq_len · global_plddt · start_index / n_term_index / c_term_index · min_seq_separation · num_contacts_eligible / contacts_emitted · num_eligible_atoms · num_events / num_distinct_atoms_mentioned · max_depth / depth1_count / depth2_count / depth3_count · frame quaternion quat_w/x/y/z + translation trans_x/y/z · truncated · num_tokens · sha1 · round · struct_cluster_id · seq_cluster_id · split · uniprot_accession / tax_id / organism_name.

sha1 = sha1 of document, so byte-equality with the MarinFold generator is a single column compare.