Rewrite the dataset card

README.md

@@ -30,7 +30,6 @@ configs:
     path: data/valid-*
   - split: test
     path: data/test-*
-
 license: cc-by-4.0
 language:
 - en
@@ -46,177 +45,125 @@ pretty_name: PLAT (Protein Language Alignment Tuples)
 task_categories:
 - feature-extraction
 - sentence-similarity
-- text-classification
+size_categories:
+- 10K<n<100K
 ---
 
 # PLAT: Protein Language Alignment Tuples
 
-`plat_data` is a curated, cluster-split dataset of protein sequences paired with structured functional annotations and natural-language descriptions. It is designed to train and evaluate cross-encoder alignment models (Vec2Vec, Mat2Mat) that map representations between different protein language models (PLMs), or between PLMs and natural-language encoders.
-
-Each row contains:
-
-- a protein `sequence` (amino-acid string),
-- an `annotation` list of integer IDs covering Gene Ontology, EC numbers, InterPro, Gene3D, cofactors, and UniProt keywords,
-- a `description` string assembled from UniProt free-text curation (function, miscellaneous notes, subcellular localization, domain).
-
-## Dataset Schema
-
-| Field         | Type        | Description |
-|---------------|-------------|-------------|
-| `sequence`    | `string`    | Canonical amino-acid sequence in single-letter IUPAC code. |
-| `annotation`  | `list[int]` | Sorted, deduplicated integer IDs decoded via the `label2id.pkl` vocabulary produced by `data/process_uniprot_av.py`. Each ID corresponds to a typed annotation key of the form `"<value>_<suffix>"` where suffix is one of `bp`, `cc`, `mf`, `ec`, `cofactor`, `ip`, `threed`, `keywords`. |
-| `description` | `string`    | Structured free-text block beginning with `"The following text describes a protein:"` and containing any present of `Function:`, `Miscellaneous:`, `Subcellular Localization:`, `Domain:` subsections. Every row is guaranteed to contain a `Function:` block (see "Preferential sampling" below). |
-| `labels`      | `int`       | Placeholder column (always `0`), added by `data/add_dummy_labels.py` so downstream `Trainer` pipelines that expect a `labels` key work without modification. |
-
-### Splits
-
-The dataset has three splits: `train`, `valid`, and `test`. Splitting is performed at the cluster level (see "Clustering and splitting" below), so no sequence in `valid` or `test` shares more than ~40% identity with any sequence in `train`.
-
-## Data Source
-
-The primary upstream source is UniProtKB, restricted to entries with manually curated Gene Ontology annotations (the "GO manual" subset). Two snapshots are used:
-
-- `uniprotkb_go_manual_2025_11_18.tsv.gz` (1,380,840 entries) for structured annotations.
-- `uniprotkb_go_manual_2025_12_02.tsv.gz` (1,380,840 entries) for free-text descriptions.
-
-Columns pulled from the UniProt TSV are:
-
-| UniProt column | Used for | Parsing |
-|----------------|----------|---------|
-| `Entry` | record key | verbatim |
-| `Sequence` | `sequence` | verbatim |
-| `EC number` | annotations | split on `;`, strip |
-| `Cofactor` | annotations | parse `Name=<x>` from each `COFACTOR:` entry |
-| `Gene Ontology (biological process)` | annotations | extract `GO:xxxxxxx` IDs |
-| `Gene Ontology (cellular component)` | annotations | extract `GO:xxxxxxx` IDs |
-| `Gene Ontology (molecular function)` | annotations | extract `GO:xxxxxxx` IDs |
-| `InterPro` | annotations | split on `;`, drop trailing empty |
-| `Gene3D` | annotations | split on `;`, drop trailing empty |
-| `Keywords` | annotations | split on `;`, strip |
-| `Function [CC]` | description | see text cleaning |
-| `Miscellaneous [CC]` | description | see text cleaning |
-| `Subcellular location [CC]` | description | see text cleaning |
-| `Domain [CC]` | description | see text cleaning |
-
-### Annotation vocabulary
-
-`data/process_uniprot_av.py` streams the UniProt TSV twice. The first pass builds a `label2id` dictionary whose keys are typed strings (e.g. `GO:0016310_bp`, `1.1.1.1_ec`, `IPR000878_ip`, `ATP-binding_keywords`, `K(+)_cofactor`). The second pass emits `camp_data.csv` with columns `[sequence, annotations]`, where `annotations` is the sorted, deduplicated list of integer IDs for that protein. The `label2id.pkl` and `id2label.pkl` files are the authoritative mapping used throughout the pipeline. The resulting table is pushed to the intermediate dataset `lhallee/camp_data_11_2025`.
-
-### Description assembly
-
-`data/process_uniprot_cc.py` builds the free-text description by concatenating the four CC columns in order (`Function`, `Miscellaneous`, `Subcellular location`, `Domain`), each under its own header. Text cleaning removes:
-
-- `(PubMed:...)` inline citations,
-- `{ECO:...}` evidence codes,
-- UniProt CC prefixes (`FUNCTION: `, `MISC: `, `SUBCELLULAR LOCATION: `, `DOMAIN: `),
-- redundant whitespace and stray punctuation artifacts from the removals above.
-
-A resulting row looks like:
-
-```
-The following text describes a protein:
-Function:
-Catalyzes the reversible conversion of methylmalonyl-CoA to succinyl-CoA...
-Subcellular Localization:
-Mitochondrial matrix...
+Every row of PLAT describes one protein three ways: its amino acid sequence, the typed functional
+annotations UniProt curators gave it, and the free text they wrote about it. The three views share no
+input, so PLAT is a test bed for asking whether the embedding spaces of protein language models,
+annotation encoders and text encoders can be aligned or translated into one another. It holds 67,402
+UniProtKB proteins with curated GO terms, one function-annotated representative per 40% identity cluster, split by
+cluster into train, validation and test. It was built for the Gleghorn Lab's study of embedding
+translation between protein representation spaces, whose frozen embeddings of these rows are in
+[GleghornLab/vec2vec_embeddings](https://huggingface.co/datasets/GleghornLab/vec2vec_embeddings).
+
+| | |
+|---|---|
+| Task | Cross-view alignment and retrieval: sequence, annotation, description |
+| Rows | 60,666 train, 3,316 valid, 3,420 test |
+| Homology | Clustered at 40% identity before the split; not disjoint under a stricter later screen (below) |
+| Source | UniProtKB entries with manually curated GO terms, release 2025_04 |
+| Licence | CC BY 4.0 |
+
+## Homology leakage
+
+**Intended: 40% identity, by cluster.** Proteins were clustered with CD-HIT at 40% identity and whole
+clusters were assigned to one split, so no sequence appears in two splits.
+
+**Measured afterwards: not homology-disjoint.** CD-HIT's greedy clustering does not guarantee that
+members of different clusters fall below the threshold, and the cluster assignment was not kept. A
+later screen with CD-HIT-2D, counting a pair when it reached 40% global identity over at least 80% of
+both sequences, found:
+
+| Held-out split | Proteins with a qualifying neighbor | Share |
+|---|---|---|
+| valid, in train | 423 of 3,316 | 12.76% |
+| test, in train | 430 of 3,420 | 12.57% |
+| test, in valid | 50 of 3,420 | 1.46% |
+| test, in train or valid | 454 of 3,420 | 13.27% |
+
+No exact sequence is shared between splits. For a strictly remote-homology evaluation, drop the 454
+test proteins with a neighbor, or recluster.
+
+## How it was built
+
+1. Read the UniProtKB entries with manually curated GO annotations, exported on 18 November 2025 (for
+   annotations) and 2 December 2025 (for free text) from release 2025_04; 1,380,840 entries each.
+2. Built the typed annotation list of each entry from eight fields: GO biological process, cellular
+   component and molecular function, EC number, cofactor, InterPro, Gene3D and UniProt keywords. Each
+   term became a typed key such as `GO:0016310_bp` or `1.1.1.1_ec`, and each key an integer id.
+3. Built the description of each entry from the Function, Miscellaneous, Subcellular location and
+   Domain comments, each under its own header, with PubMed citations, ECO evidence codes and the
+   comment prefixes removed.
+4. Removed exact duplicate sequences.
+5. Clustered the sequences with CD-HIT at 40% identity (word size 5) and assigned whole clusters to
+   train, validation and test in the proportions 90, 5 and 5, after a shuffle.
+6. From each cluster, kept the members that had both annotations and a description containing
+   Function text, and drew one of them at random (seed 42). A cluster with no such member was dropped,
+   so PLAT leans toward well characterized families.
+7. Added a constant `labels` column so generic training loops that expect one run unchanged.
+
+## Columns
+
+| Column | Type | What it is |
+|---|---|---|
+| `sequence` | string | Canonical amino acid sequence, single-letter code; median length 403, longest 8,923 |
+| `annotation` | list of int | Sorted, deduplicated typed annotation ids; median 18 per protein, 56,868 distinct ids in use |
+| `description` | string | UniProt curator text under `Function:` and, when present, `Miscellaneous:`, `Subcellular Localization:` and `Domain:` headers; median 339 characters |
+| `labels` | int | Always 0; a placeholder, not a label |
+
+The integer ids index the Annotation Vocabulary (Hallee et al., 2024), here built from this UniProt
+release with about 88,000 typed keys across the eight annotation types. The id to key mapping is not
+yet published with this dataset.
+
+## Example rows
+
+From the test split, sequences and lists shortened.
+
+| sequence | annotation | description |
+|---|---|---|
+| `MDKKEILKEFSADPDKYYNV...` (889 aa) | 31 ids: 6315, 8144, 24978, ... | Function: Catalyzes the attachment of alanine to tRNA(Ala) in a two-step reaction... |
+| `MSHQTGIKGNVELISFFGKC...` (344 aa) | 18 ids: 9957, 10213, 11596, ... | Function: Actin-binding protein involved in motile and morphological processes. Inhibits actin polymerization... Subcellular Localization: Cytoplasm, cell cortex. Cytoplasm, cytoskeleton. |
+| `MTISMISPSSIRLSSDKRDS...` (410 aa) | 23 ids: 9619, 15796, 18993, ... | Function: Promotes excretion of zinc from intestinal cells into the intestinal lumen in response to increased dietary zinc... |
+
+```python
+from datasets import load_dataset
+
+plat = load_dataset("GleghornLab/plat_data")
+row = plat["test"][0]
+row["sequence"], row["annotation"], row["description"]
 ```
 
-## Build Pipeline
-
-The full pipeline is four stages. All scripts live under [data/](data/) in the source repo.
-
-### Stage 1: UniProt ingestion and annotation curation
-
-- Input: `uniprotkb_go_manual_2025_11_18.tsv.gz` and `uniprotkb_go_manual_2025_12_02.tsv.gz`.
-- [data/process_uniprot_av.py](data/process_uniprot_av.py) produces `camp_data.csv` (sequence + integer annotation list) and the `label2id.pkl` / `id2label.pkl` vocabulary files.
-- [data/process_uniprot_cc.py](data/process_uniprot_cc.py) produces `seq_descriptions.tsv` (entry + sequence + structured description).
+## What a model trained on it may say
 
-The sequence-keyed dictionaries `seq_annotation_dict.pkl` and `seq_description_dict.pkl` are later used by the final assembly step to look up annotations and descriptions by exact sequence match.
+A model trained on PLAT maps one view of a protein to another: a sequence embedding to an annotation or
+description embedding, or back. A good retrieval score says the views of the same protein can be found
+from each other among held-out proteins. It does not say the model understands protein function: a
+sequence-only control can retrieve well through shared input alone, and about one held-out protein in
+eight has a close relative in train. Every protein has curated GO terms and Function text, so the
+dataset says little about poorly characterized proteins.
 
-The annotation table is pushed to [lhallee/camp_data_11_2025](https://huggingface.co/datasets/lhallee/camp_data_11_2025) as an intermediate artifact.
+## Sources and licence
 
-### Stage 2: Deduplication
-
-Before clustering, exact-duplicate sequences are removed with `drop_duplicates(subset=['sequence'])` (see [data/build_dataset.py](data/build_dataset.py)).
-
-### Stage 3: Clustering and splitting
-
-Clustering and split assignment are performed by [data/build_dataset.py](data/build_dataset.py):
-
-- Sequences are written to a FASTA file with integer IDs.
-- CD-HIT is invoked inside a Docker container built from the official CD-HIT Dockerfile.
-- **Identity threshold: 0.4** (40% sequence identity), word size `-n 5`. This is the threshold actually used for the published PLAT release, as recorded by the cluster filename `output_lhallee_camp_data_11_2025_0.4.clstr`.
-- The `.clstr` output is parsed into `cluster_dict: {cluster_id -> [seq_ids]}` and `id_seq_dict: {seq_id -> sequence}`.
-- Clusters are shuffled and partitioned:
-  - **5% valid clusters, 5% test clusters, 90% train clusters.**
-- The four working pickles (`cluster_dict.pkl`, `id_seq_dict.pkl`, `final_cluster_dict.pkl`, `seq_annotation_dict.pkl`) are uploaded to an auxiliary private repo for reproducibility.
-
-Because splits are assigned by cluster, any two sequences in different splits are less than ~40% identical under CD-HIT's greedy incremental clustering, which gives a realistic generalization benchmark for PLM alignment.
-
-### Stage 4: PLAT assembly (preferential sampling)
-
-[data/build_plat_data.py](data/build_plat_data.py) produces the final dataset:
-
-1. For each cluster in each split, gather all member sequences.
-2. Keep only sequences for which both an `annotation` and a `description` exist.
-3. Keep only the subset whose `description` contains the literal substring `"Function:"` (that is, sequences with at least a UniProt `Function [CC]` annotation after text cleaning).
-4. Pick one sequence uniformly at random from that subset (seed `42`).
-5. If no candidate survives, **the cluster is skipped entirely**. This biases PLAT toward clusters containing at least one well-characterized member.
-6. The resulting records (`sequence`, `annotation`, `description`) are assembled into a `DatasetDict(train, valid, test)` and pushed to `lhallee/plat_data`.
-
-A final `labels` column of constant `0` is added by [data/add_dummy_labels.py](data/add_dummy_labels.py) for compatibility with generic `Trainer` pipelines.
-
-## Reproducing the Build
-
-```bash
-# 1. Parse UniProt TSV into annotations + vocabulary
-py -m data.process_uniprot_av
-# -> camp_data.csv, label2id.pkl, id2label.pkl
-# -> pushes lhallee/camp_data_11_2025
-
-# 2. Parse UniProt TSV into descriptions
-py -m data.process_uniprot_cc
-# -> seq_descriptions.tsv
-
-# 3. Cluster with CD-HIT and split by cluster (requires Docker)
-py -m data.build_dataset \
-    --hf_token <token> \
-    --dataset_path lhallee/camp_data_11_2025 \
-    --similarity_threshold 0.4 \
-    --n 5 \
-    --valid_percentage 0.05 \
-    --test_percentage 0.05
-
-# 4. Assemble PLAT with preferential Function-aware sampling
-py -m data.build_plat_data \
-    --hf_token <token> \
-    --repo_name lhallee/plat_data \
-    --seed 42
-```
-
-## Intended Use
-
-PLAT was constructed to train and evaluate representation-alignment models:
-
-- **Vec2Vec**: align pooled (mean + variance) embeddings from one PLM to another, or from a PLM to a natural-language encoder of the `description` field.
-- **Mat2Mat**: align full residue-by-residue embedding matrices between two PLMs.
-- **Cross-modal retrieval**: use the `description` column with a text encoder (for example ModernBERT or GPT-OSS) and the `sequence` column with a PLM (for example ESMC-600), then evaluate whether aligned embeddings retrieve each other.
-
-Typical consumption patterns in the source repo read columns as `(sequence, sequence)`, `(sequence, description)`, or `(description, description)` pairs.
-
-## Licensing and Attribution
-
-All sequences and annotations are derived from UniProtKB and are redistributed under the UniProt terms of use (Creative Commons Attribution 4.0, `CC BY 4.0`). If you use PLAT, please also cite UniProt:
+All sequences, annotations and descriptions come from
+[UniProtKB](https://www.uniprot.org), redistributed under its
+[CC BY 4.0 licence](https://www.uniprot.org/help/license). The annotation types draw on the Gene
+Ontology, the IUBMB Enzyme Commission, InterPro, CATH-Gene3D and the UniProt keywords. If you use PLAT,
+please cite UniProt:
 
 > The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2025. *Nucleic Acids Research*, 2025.
 
-And the dependent databases surfaced in the annotation vocabulary: Gene Ontology, Enzyme Commission (IUBMB), InterPro, Gene3D, and the UniProt Keyword ontology.
-
-## Known Limitations
+Limits worth knowing: only entries with manually curated GO terms were read, so taxa that curators have
+studied little are thin here; each cluster contributes one protein, so paralogs and species variants are absent; the
+descriptions lose their citations and evidence codes; and the data reflect one UniProt release.
 
-- **Manual-annotation bias.** Only UniProt entries with manually curated GO terms are considered, and only clusters containing at least one member with a `Function [CC]` block appear in the final dataset. Taxa and protein families that are underrepresented in Swiss-Prot are correspondingly underrepresented here.
-- **One representative per cluster.** Within-cluster diversity (paralogs, species variants) is not preserved: each cluster contributes a single randomly chosen, function-annotated representative.
-- **Descriptions are post-processed.** PubMed citations, ECO evidence codes, and the CC-prefix tags are stripped, which simplifies downstream tokenization but loses provenance information relative to the raw UniProt text.
-- **Integer annotation labels are only meaningful via `label2id.pkl`.** The integer space mixes eight annotation types; training code that treats it as a single flat multi-label target should be aware that different IDs correspond to semantically different taxonomies.
-- **Snapshot-frozen.** PLAT is built from a fixed UniProt release (manual-GO subset, November-December 2025) and will not reflect later curation updates until rebuilt.
+## Files
 
+| File | What it is |
+|---|---|
+| `data/train-00000-of-00001.parquet` | 60,666 training rows |
+| `data/valid-00000-of-00001.parquet` | 3,316 validation rows |
+| `data/test-00000-of-00001.parquet` | 3,420 test rows |