Datasets:

lsieben
/

climb-pretrain-data

text stringlengths 1 372
B.CC(=O)OC1CN2CCC1CC2
Br.Br.N=C(N)SCc1cn2ccccc2n1
Br.Brc1ccc(C2=CSC3=NCCN23)cc1
Br.C1=C(c2ccccc2)N2CCN=C2S1
Br.C=CC1CN2CCC1CC2C(O)c1ccnc2ccc(OC)cc12
Br.C=CCN1CC2CN=NC2(C#N)C1
Br.C=CCn1c(-c2ccc(Cl)cc2)csc1=Nc1ccccc1OC
Br.CC(C)(CCCCC(C)(C)NC1=NCCO1)NC1=NCCO1
Br.CC(CN(C)C)C(=O)c1ccccc1
Br.CC(Cc1ccc(O)cc1)NCC(O)c1cc(O)cc(O)c1
Br.CC(N)Cc1ccc(O)cc1
Br.CC(NCCCNC(C)C(=O)O)C(=O)O
Br.CC1=CC(C)=Nc2ccccc2N1
Br.CC1CCCCN1CC(=O)c1ccc2c(c1)Sc1ccccc1O2
Br.CCC1=C(CC2NCCc3ccccc32)CC2c3cc(OC)c(OC)cc3CCN2C1
Br.CCCC(NCCCCCCNC(CCC)C(=O)O)C(=O)O
Br.CCCCCCCCN1C=CN(C)C1
Br.CCCCCCCN=c1ssc(=O)n1CCCCCCC
Br.CCCCN1C=CN(C)C1
Br.CCCN1C=CN(C)C1
Br.CCCN=c1ssc(=O)n1CCC
Br.CCN
Br.CCN(CC)CCC(=O)c1ccccc1
Br.CCN(CC)CCN(C)CCc1ccc(Br)cc1
Br.CCN(CC)CCSc1c2ccccc2nc2cc(NCC(=O)Nc3ccccc3-c3ccccc3NC(=O)CNc3ccc4c(SCCN(CC)CC)c5ccccc5nc4c3)ccc12
Br.CCN1C=CN(C)C1C
Br.CCN1CCC(O)(c2ccc(C)cc2)C(C(=O)c2ccc(C)cc2)C1
Br.CCN1CCC(O)(c2ccc(Cl)c(Cl)c2)C(C(=O)c2ccc(Cl)c(Cl)c2)C1
Br.CCN1CCC(O)(c2ccc(Cl)cc2)C(C(=O)c2ccc(Cl)cc2)C1
Br.CCN1CCC(O)(c2ccc(OC)cc2)C(C(=O)c2ccc(OC)cc2)C1
Br.CCN1CCC(O)(c2ccccc2)C(C(=O)c2ccccc2)C1
Br.CCOC(=O)C(Cc1ccc(O)cc1)NC(=O)C(N)CSSCC(N)C(=O)NC(Cc1ccc(O)cc1)C(=O)OCC
Br.CCOC(=O)c1c(N2CCCC2)nsc1Nc1ccccc1
Br.CCOC(CN=c1nc(N(C)C)ss1)OCC
Br.CCOC1CN(C)C(=Nc2cccc3ccccc23)S1
Br.CCc1cccc(CC)c1Nc1nc(-c2ccccc2-c2ccccc2)cs1
Br.CN(C)CC1CCCCCC(CN(C)C)C1=O
Br.CN(C)CCCC1(c2ccc(F)cc2)OCc2cc(C#N)ccc21
Br.CN(C)CCCSc1c2ccccc2nc2cc(NCC(=O)Nc3ccccc3-c3ccccc3NC(=O)CNc3ccc4c(SCCCN(C)C)c5ccccc5nc4c3)ccc12
Br.CN(C)CCNc1ncnc2nc(-c3ccccc3)c(-c3ccccc3)nc12
Br.CN(C)CCSc1nccc(-c2ccc(-c3ccnc(SCCN(C)C)n3)s2)n1
Br.CN(C)c1nc(=NCc2ccccc2)ss1
Br.CN(C)c1nc(=NCc2ccccn2)ss1
Br.CN(C)c1nc(=NCc2cccnc2)ss1
Br.CN(C)c1nc(=NCc2cccs2)ss1
Br.CN(C)c1nc(=NCc2ccncc2)ss1
Br.CN(C)c1nc(=Nc2ccc3c(c2)OCO3)ss1
Br.CN(C)c1nc(=Nc2cccc(Cl)c2)ss1
Br.CN(C1=NCCO1)c1cccc2ccccc12
Br.CN(CCc1ccc(Br)cc1)CCN1CCCCC1
Br.CN1C(C)(C)C(Br)C(=O)C(Br)C1(C)C
Br.CN1C(C)(C)CC(=O)C(Br)C1(C)C
Br.CN1C2CCC1CC(OC(=O)C(O)c1ccccc1)C2
Br.CN1CCc2cc(Cl)c(O)cc2[C@H]2c3ccccc3CC[C@@H]21
Br.CN1[C@@H]2CC(OC(=O)[C@H](CO)c3ccccc3)C[C@H]1[C@@H]1O[C@@H]12
Br.CN1[C@@H]2CC(OC(=O)[C@H](CO)c3ccccc3)C[C@H]1[C@@H]1O[C@@H]12.O
Br.CN1[C@@H]2CC(OC(=O)[C@H](CO)c3ccccc3)C[C@H]1[C@@H]1O[C@@H]12.O.O.O
Br.CN=C(Nc1cccc(C(F)(F)F)c1)SCc1ccc(N=C=S)cc1
Br.COC(=O)C1=CCCN(C)C1
Br.COC1CN(C)C(=Nc2cccc3ccccc23)S1
Br.COc1c2c(cc3c1OCO3)CCN(C)C2
Br.COc1cc(CC(C)N)c(OC)cc1Br
Br.COc1cc(CN=c2nc(N(C)C)ss2)cc(OC)c1OC
Br.COc1cc(CNC2=NCCC2)cc(OC)c1OC
Br.COc1ccc(-c2cn3ccsc3n2)cc1
Br.COc1ccc(-c2csc(=Nc3ccccc3)n2CC(O)c2ccc([N+](=O)[O-])cc2)cc1
Br.COc1ccc(NC2=NCCO2)c2ccccc12
Br.COc1ccc2c(c1)[C@@]13CCCCC1[C@@H](C2)N(C)CC3
Br.COc1ccc2c(c1)[C@@]13CCCCC1[C@@H](C2)N(C)CC3.O
Br.COc1ccc2c(c1)[C@]13CCCC[C@@H]1[C@H](C2)N(C)CC3.O
Br.COc1ccc2c3c1OC1CC(O)C=CC31CCN(C)C2
Br.COc1ccc2c3c1O[C@H]1C[C@@H](O)C=C[C@@]31CCN(C)C2
Br.C[C@@H]1[C@@H]2Cc3ccc(O)cc3[C@@]1(C)CCN2CCc1ccccc1
Br.C[C@H]1[C@H]2Cc3ccc(O)cc3[C@]1(C)CCN2CCc1ccccc1
Br.C[N+]12CCCN(CC1)CC2.O=[N+]([O-])c1cc([N+](=O)[O-])c(O)c([N+](=O)[O-])c1.[Br-]
Br.Cc1c(O)c(O)c(C)c(CCN)c1O
Br.Cc1cc(C2=CSC3=NCCN23)ccc1Cl
Br.Cc1ccc(-c2csc(=Nc3ccccc3)n2NC(=O)c2ccncc2)cc1
Br.Cc1ccc(C2=CSC3=NCCN23)cc1Cl
Br.Cc1cn2cc(-c3cccs3)nc2s1
Br.Cc1nc(-[n+]2nc(-c3ccncc3)nn2-c2ccccc2)sc1C
Br.Cc1nc(Nc2ccc(Cl)cc2Cl)sc1C(=O)Nc1ccccc1
Br.Clc1ccc(C2=CSC3=NCCN23)cc1
Br.Clc1ccc(Cl)c(-c2csc(-c3cccnc3)n2)c1
Br.Cn1c(=O)c2c(ncn2C)n(C)c1=O
Br.N#CC(C#N)=C(C#N)C#N.c1ccncc1
Br.N=C(N)SCCCN
Br.N=C(N)SCc1ccc(N=C=S)cc1
Br.N=C(N)SCc1cn2ccccc2n1
Br.N=c1sc(N2CCOCC2)nn1CC(=O)c1ccccc1
Br.NC(=O)C(=CNc1nc(-c2ccc(Br)cc2)cs1)C(N)=O
Br.NC(=O)C(c1ccccc1)(c1ccccc1)[C@@H]1CCN(CCc2ccc3c(c2)CCO3)C1
Br.NC(=O)NC1=NCCC(=O)N1CCc1c[nH]c2ccccc12
Br.NC(CNC(=O)CBr)C(=O)O
Br.NC1=CN2c3ccccc3N(CC(=O)O)C2C=C1
Br.NC1CCCCC1
Br.NC1NCCC(=O)N1CCc1c[nH]c2ccccc12
Br.NC1OC(COC(=O)c2ccccc2)C(O)C(OC(=O)c2ccccc2)C1OC(=O)c1ccccc1
Br.NCCC1([N+](=O)[O-])CCNCC1
Br.NCCCCNOCCCN

End of preview. Expand in Data Studio

CLIMB pretraining data

The corpora, descriptor targets and auxiliary tables used to pretrain the CLIMB encoders.

Layout

tokenized_sources/pubchem_filtered/            source SMILES after filtering
tokenized_sources/pubchem_filtered_*_pkl/      tokenized pretraining corpora (see below)
tokenized_sources/pubchem_descriptors/         217 RDKit descriptor targets, sharded
tokenized/supervised_wide_parquet/             supervised fine-tuning table
tokenizer/, tokenizers_vocab/                  the vocab-1000 tokenizer and the vocab-sweep variants
configs/descriptor_stats.json                  descriptor names and normalisation statistics

Pretraining corpora

Corpus	Contents
`pubchem_filtered_tokenized_pkl`	the real PubChem SMILES corpus
`pubchem_filtered_bigram_pkl`	sequences resampled from the corpus bigram statistics: local adjacency only
`pubchem_filtered_unigram_pkl`	sequences resampled from the corpus unigram marginal: no sequential structure
`pubchem_filtered_wiki_pkl`	English Wikipedia text, tokenized with the same tokenizer: no chemistry

The token-shuffled control is applied as a training-time transform of the real corpus and has no separate artifact.

The 124M-molecule RDKit-canonical corpus is not re-hosted. It derives from hheiden/PubChem-124M-SMILES-SELFIES-InChI-IUPAC; scripts/download_pubchem_full.sh rebuilds the exact copy used here.

Leakage

Molecules overlapping the downstream evaluation sets are recorded in the blocklist and excluded; the audit procedure is described in METHODS.md.

Code: github.com/leifsieben/CLIMB
Encoders: lsieben/climb-encoders
Results: lsieben/climb-results
Pretraining data: lsieben/climb-pretrain-data

Citation

@misc{climb2026,
  title  = {Does Pretraining Teach Chemical Language Models Chemistry?},
  author = {Sieben, Leif and Zimmermann, Yoel},
  year   = {2026},
  note   = {Preprint, arXiv},
  url    = {https://github.com/leifsieben/CLIMB}
}

License

CC-BY-4.0.

Downloads last month: -