Upload README.md with huggingface_hub

README.md

@@ -6,6 +6,7 @@ tags:
   - host-guest
   - molecular-recognition
   - macrocyclic
+  - continued-pretraining
 size_categories:
   - 100K<n<1M
 configs:
@@ -19,20 +20,15 @@ configs:
 
 # EU-PMC Supramolecular Chemistry Corpus
 
-A collection of titles and abstracts from Europe PMC for supramolecular chemistry research.
-Two splits are provided:
+A curated collection of titles and abstracts from Europe PMC for supramolecular chemistry
+research, intended for continued pre-training of language models in the chemistry domain.
 
-- **raw** (~420K articles): Unfiltered superset matched by 19 keyword-based search topics on Europe PMC
-  (supramolecular, host-guest, self-assembly, molecular recognition, cyclodextrin, crown ether,
-  calixarene, cucurbituril, pillararene, rotaxane, dendrimer, MOF, non-covalent, molecular cage,
-  inclusion complex, supramolecular polymer, metallosupramolecular, macrocyclic chemistry, and
-  binding-affinity-host queries). Includes biology/medical contamination from broad MeSH-tag matches.
+## Dataset Summary
 
-- **filtered** (~134K articles): High-precision subset where the title or abstract literally contains
-  one of ~440 supramolecular-specific keywords (e.g., supramolecular, cyclodextrin, [n]rotaxane,
-  cucurbit[n]uril, calix[n]arene, pillar[n]arene, MOF, COF, cage, foldamer, etc.). Two-tier negative
-  filter rejects clear biology contamination (SARS-CoV-2, T-cell, etc.). Suitable for continued
-  pre-training of language models on supramolecular chemistry domain.
+| Split | Articles | Description |
+|---|---:|---|
+| **raw** | 420,950 | Unfiltered superset matched by 19 Europe PMC search topics |
+| **filtered** | 133,867 | High-precision subset via tiered keyword filtering |
 
 ## Columns
 
@@ -45,19 +41,205 @@ Two splits are provided:
 ```python
 from datasets import load_dataset
 
-# High-precision filtered set (recommended for continued pre-training)
-ds = load_dataset("SupraBench/EU-PMC", split="filtered")
+# Recommended for continued pre-training (high precision)
+ds = load_dataset("SupraBench/EU-PMC", split="filtered")  # 133,867 articles
 
-# Full superset
-ds_raw = load_dataset("SupraBench/EU-PMC", split="raw")
+# Full superset (broader recall, more contamination)
+ds_raw = load_dataset("SupraBench/EU-PMC", split="raw")  # 420,950 articles
 ```
 
+---
+
+## Construction Pipeline
+
+### Stage 1: Source Collection
+
+Articles were collected from **Europe PMC** (https://europepmc.org/), the European mirror of
+PubMed Central, which indexes 40+ million biomedical and life-sciences citations.
+
+We harvested the **complete Europe PMC abstracts corpus** (1900-2026, ~225 GB JSON) using the
+REST API search endpoint with cursor-based pagination, organized by publication year.
+
+### Stage 2: Topical Subset Selection (→ raw split)
+
+We ran **19 separate Europe PMC search queries**, each targeting a supramolecular chemistry
+sub-topic. Each query was paginated through the REST API (1,000 results per page) and
+deduplicated by PMID across all queries.
+
+| Topic | Query |
+|---|---|
+| supramolecular | `supramolecular` |
+| host-guest | `"host-guest" OR "host guest chemistry" OR "host-guest binding"` |
+| self-assembly | `"self-assembly" OR "self assembly" AND (molecular OR supramolecular)` |
+| molecular-recognition | `"molecular recognition"` |
+| cyclodextrin | `cyclodextrin OR cyclodextrins` |
+| crown-ether | `"crown ether" OR "crown ethers"` |
+| calixarene | `calixarene OR calixarenes` |
+| cucurbituril | `cucurbituril OR cucurbiturils` |
+| pillararene | `pillararene OR pillararenes` |
+| rotaxane / catenane | `rotaxane OR rotaxanes OR catenane OR catenanes` |
+| dendrimer | `dendrimer OR dendrimers` |
+| metal-organic-framework | `"metal-organic framework" OR "metal organic framework" OR MOF` |
+| non-covalent | `"non-covalent interaction" OR "noncovalent interaction" OR "supramolecular interaction"` |
+| molecular-cage | `"molecular cage" OR "nanocage" OR "supramolecular cage"` |
+| binding-affinity-host | `"binding affinity" AND (host OR macrocyclic OR supramolecular)` |
+| inclusion-complex | `"inclusion complex" OR "inclusion complexes"` |
+| supramolecular-polymer | `"supramolecular polymer" OR "supramolecular polymers"` |
+| metallosupramolecular | `metallosupramolecular OR metallo-supramolecular` |
+| macrocyclic-chemistry | `"macrocyclic chemistry" OR "macrocycle" OR "macrocycles"` |
+
+**Pre-cleaning:** retained only articles with non-empty abstract (length ≥ 50 characters)
+and unique PMID.
+
+This stage produces the **raw split** with **420,950 unique articles**.
+
+### Stage 3: Tiered Keyword Filtering (→ filtered split)
+
+The raw split contains substantial biology/medical contamination because Europe PMC's search
+matches against **MeSH headings** and **keyword lists** — not just title/abstract text. For
+example, the `binding-affinity-host` query matched ~70K papers about pathogens binding to
+*host cells* rather than supramolecular host-guest chemistry.
+
+To produce the high-precision **filtered split**, each article was assigned to one of five
+clusters based on regex matching against its title and abstract.
+
+#### Keyword Lists
+
+The filter uses three regex patterns compiled from curated keyword lists:
+
+**Positive keywords (≈440 terms across 5 sub-lists):** terms that, when matched in title or
+abstract, indicate the article is supramolecular chemistry.
+
+- `POSITIVE_TIER1` (~165 core terms): `supramolecular`, `host-guest`, `mechanically interlocked`,
+  `[2]rotaxane`/`[3]rotaxane`/etc., `cyclophane`, `cryptand`, `cavitand`, `foldamer`,
+  `cyclodextrin`/`β-CD`/`α-CD`/`γ-CD`/`HP-β-CD`/`HPBCD`, `cucurbit[5–10]uril`/`CB[5–10]`,
+  `calix[4]arene`/`calix[6]arene`/`calix[8]arene`/`thiacalix`, `resorcinarene`,
+  `pillar[5]arene`/`pillar[6]arene`, `crown ether`/`18-crown-6`/`15-crown-5`/etc.,
+  `metal-organic framework`/`COF`/`ZIF-8`/`HKUST-1`/`UiO-66`/named MOFs, `inclusion complex`,
+  `coordination cage`, `molecular cage`/`molecular capsule`, etc.
+- `EXTENDED_POSITIVE` (~75 terms): qualified self-assembly phrases (`molecular self-assembly`,
+  `peptide self-assembly`, `templated self-assembly`, etc.), peptide nanostructures,
+  small-molecule gelators, polymer micelles, metallacycles/metallacages, porous organic cages,
+  hydrogen-bonded frameworks, fullerene host-guest, daisy chains, sub-component self-assembly.
+- `EXTENDED_POSITIVE_V2` (~70 terms): ionophores (`valinomycin`, `nigericin`), hierarchical
+  assembly, polymer self-assembly variants, nanogels, crystal engineering (`supramolecular
+  synthon`, `cocrystal`), DNA nanotechnology (`DNA origami`, `G-quadruplex`), discrete
+  containers, stimuli-responsive supramolecular, aromatic building blocks (`perylene diimide`,
+  `naphthalenediimide`).
+- `EXTENDED_POSITIVE_V3` (~38 terms): bare core terms missing from earlier tiers — `rotaxane`,
+  `catenane`, `macrocyclic`, `macrocycle`, `metallacycle`, `metallosupramolecular`, crown
+  variants, cyclophane variants, molecular box.
+- `EXTENDED_POSITIVE_V4` (~90 terms): fullerenes (`C60`, `C70`, `buckminsterfullerene`),
+  carcerand/hemicarcerand, single-molecule magnets, polyoxometalates, nanographene,
+  molecular springs/muscles/ratchets/pumps, aromatic recognition (`cation-π`),
+  coordination polymers (1D/2D/3D), specific cage stoichiometries (`M2L4`, `M4L6`,
+  `Pd2L4`), boroxine, imine cages, Schiff base macrocycles, chiral self-assembly.
+
+**Negative keywords:**
+
+- `NEGATIVE_TITLE_HARD` (~140 terms): hard-reject if found in title regardless of other
+  signals — `SARS-CoV-2`, `COVID-19`, `T-cell receptor`, `T-cell activation`, `HIV-1
+  infection`, `bacterial infection`, `bacterial pathogen`, `Staphylococcus aureus`,
+  `gene expression regulation`, `RNA splicing`, `genome-wide association`, `mouse model
+  of`, `clinical trial`, `randomized controlled`, etc.
+- `NEGATIVE_TITLE_CONDITIONAL` (~50 terms): reject only if title has no supramolecular
+  anchor — `vaccine adjuvant`, `monoclonal antibody`, `cancer cells`, `apoptosis`,
+  `gene expression`, `biofilm`, `inflammation`, `diabetes`, etc.
+
+#### Cluster Assignment
+
+Each article is assigned to exactly one cluster, in priority order:
+
+| Cluster | Rule | Action |
+|---|---|---|
+| **A** | Title contains a positive keyword | KEEP |
+| **B** | Abstract contains a positive keyword (and Cluster A didn't fire) | KEEP |
+| **C** | Title contains a hard-reject negative keyword (and Cluster A/B didn't fire) | REJECT |
+| **D** | Title contains a conditional-reject negative keyword and no positive anchor | REJECT |
+| **E** | None of the above | EXCLUDED from filtered split |
+
+**Note:** Positive signals override negative signals — a paper titled "Cucurbit[7]uril
+binding in cancer cells" (positive: `cucurbit[7]uril`, negative: `cancer cells`) lands in
+Cluster A and is kept.
+
+#### Cluster Distribution
+
+| Cluster | Articles | Output |
+|---|---:|---|
+| A — Title positive | 76,281 | filtered split |
+| B — Abstract positive | 57,586 | filtered split |
+| C — Hard reject | 27,481 | dropped |
+| D — Conditional reject | 32,633 | dropped |
+| E — No signal (mostly biology contamination) | 226,969 | excluded |
+| **Total** | **420,950** | — |
+| **filtered split (A + B)** | **133,867** | uploaded |
+
+Cluster E contains primarily biology/medical/materials papers that matched a topical search
+via Europe PMC's MeSH-tag indexing but never explicitly mention supramolecular vocabulary
+in their title or abstract. Manual inspection of random samples showed ~85-90% of Cluster E
+is non-supramolecular content (cell biology, microbiology, clinical research, plant
+biology, etc.), so the entire cluster is excluded from the filtered split.
+
+### Stage 4: Output Format
+
+Both splits are stored as **Parquet** files with `zstd` compression in the `data/` directory:
+
+- `data/raw-NNNNN-of-00009.parquet` — 9 shards, ~50K rows each, 230 MB total
+- `data/filtered-NNNNN-of-00003.parquet` — 3 shards, ~45K rows each, 70 MB total
+
+---
+
+## Quality Notes
+
+### Strengths
+
+- **Complete temporal coverage** (1900-2026) — includes historical foundational works
+  (e.g., Cram, Pedersen, Lehn) and modern advances
+- **Both open-access and paywalled abstracts** — PMID-keyed dataset captures abstracts
+  even for subscription-only journals
+- **High-precision filtered split** — manual sampling showed >85% of Cluster A and >70% of
+  Cluster B samples are clearly supramolecular chemistry
+- **Transparent filter rules** — all keywords are documented and reproducible
+
+### Limitations
+
+- **No full-text content** — only titles and abstracts are included
+- **No structural data** — molecular structures, binding constants, and figures are not
+  extracted (would require full-text mining)
+- **English-only** — Europe PMC primarily indexes English-language abstracts
+- **Filter is title/abstract-bound** — articles that discuss supramolecular concepts using
+  uncommon vocabulary may be missed (Cluster E)
+- **Some biochemistry overlap in filtered split** — terms like `supramolecular complex`
+  appear in plant photosystem papers; `hierarchical self-assembly` matches some
+  collagen/cytoskeleton biochemistry. Estimated false-positive rate in filtered split: ~10-15%.
+
+### Validation
+
+Multiple random samples from each cluster were manually inspected. Two examples were
+cross-checked against the original PubMed Central source to confirm the title and abstract
+text matches the published paper:
+
+- PMID 36686943: *"Solution- and gas-phase study of binding of ammonium and bisammonium
+  hydrocarbons to oxacalix[4]arene carboxylate"* — match confirmed
+- PMID 18690725: *"Ring-opening polymerization of cyclic esters by cyclodextrins"* — match
+  confirmed
+
+---
+
 ## Source
 
-Articles are sourced from Europe PMC (https://europepmc.org/) under their respective open-access
-licenses. See https://europepmc.org/Copyright for the licensing of individual articles.
+Articles are sourced from **Europe PMC** (https://europepmc.org/) under their respective
+open-access or restricted licenses. See https://europepmc.org/Copyright for the licensing
+terms governing individual articles.
+
+This corpus is a derived work that includes only article titles, abstracts, and PMIDs.
+Users should consult the original article for full content and licensing.
 
 ## Citation
 
-The Europe PMC Consortium. Europe PMC: a full-text literature database for the life sciences and
-platform for innovation. *Nucleic Acids Research*, 2014.
+If you use this dataset, please cite Europe PMC:
+
+> The Europe PMC Consortium. Europe PMC: a full-text literature database for the life
+> sciences and platform for innovation. *Nucleic Acids Research*, 2014.
+
+And acknowledge the SupraBench / EU-PMC corpus by linking to this dataset.