Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Improving Chemical Understanding of LLMs via SMILES Parsing

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Pretraining on deterministic SMILES parsing tasks gives LLMs a transferable grasp of molecular structure.

desk verdict A useful incremental pretraining recipe for chemistry LLMs; the transfer claim holds up only after a molecule-overlap check between ZINC250k and Mol-Instructions. read the letter →

arxiv 2505.16340 v1 pith:K7YLIDBM submitted 2025-05-22 cs.LG

Yunhui Jang , Jaehyung Kim , Sungsoo Ahn This is my paper
classification cs.LG
keywords SMILESparsingmolecularpretrainingLLMsforchemistrygenerationcurriculumlearningdatapruningstructuralunderstandingreactionprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the bottleneck for large language models in chemistry is not missing chemical knowledge but the inability to read the molecular graph out of a SMILES string. It proposes five deterministic parsing tasks, labeled automatically by cheminformatics software, and uses them to pretrain open-weight 7-8B models before fine-tuning on downstream tasks. The experiments report consistent gains on reaction-related generation benchmarks, with parsing-pretrained models reaching state-of-the-art or competitive results while never seeing generation examples during pretraining. The payoff, if the claim holds, is that cheap and unlimited structural supervision can substitute for expensive task-specific chemical data.

What carries the argument

The load-bearing mechanism is SMILES parsing reformulated as a suite of deterministic tasks: three subgraph-matching tasks (functional group presence, ring-size counting, longest carbon chain) and two global-graph tasks (canonical SMILES generation and fragment assembly). Each has a unique answer that an open-source cheminformatics toolkit can compute for any valid molecule, making the supervision unambiguous and scalable. A per-task difficulty heuristic (counts of functional groups, rings, branches, and SMILES length) drives data pruning toward mid-difficulty molecules and orders examples from easy to hard in a curriculum, and the whole suite is applied as a supervised pretraining stage before downstream fine-tuning.

What would settle it

Compute the canonical-SMILES overlap between the 250K pretraining molecules and the benchmark test molecules; if a substantial fraction of test products or reactants already appeared in pretraining, the reported gains could come from memorization rather than parsing transfer, and a scaffold-disjoint retest would settle the question.

Watch

Extended reading notes

Core claim

CLEANMOL establishes that pretraining an LLM on unambiguous, automatically labeled SMILES parsing tasks—functional group matching, ring counting, carbon chain length measurement, SMILES canonicalization, and fragment assembly—substantially improves downstream molecular generation. On the Mol-Instructions benchmark, the parsing-pretrained LLaMA-3.1-8B and Qwen2.5-7B models match or beat strong baselines on retrosynthesis, reagent prediction, and forward reaction prediction, with consistent gains over the same backbones fine-tuned without the parsing stage. The authors interpret this as evidence that clean structural supervision transfers to generation tasks even without task-specific pretraining data.

Load-bearing premise

The downstream gains are interpreted as transfer of structural understanding rather than memorization, which requires that the benchmark's test molecules were effectively unseen during the 250K-molecule parsing pretraining; the paper follows the benchmark splits but does not verify overlap or deduplication.

Editorial extensions

If this is right

  • General-purpose LLMs that score below 60 percent on most parsing tasks can be taught to parse accurately with supervised fine-tuning, closing a basic competence gap.
  • Multi-task parsing training outperforms single-task training on four of five tasks, indicating that the structural skills are shared rather than task-specific.
  • Parsing pretraining gives consistent downstream gains on retrosynthesis, reagent prediction, and forward reaction prediction for both tested 7-8B backbones.
  • Selecting mid-difficulty molecules and ordering them by difficulty improves parsing accuracy over random or SMILES-length-based selection.
  • Scaling from 10K to 50K molecules per task monotonically improves parsing and retrosynthesis performance, supporting scalability of the approach.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the transfer is genuine, SMILES parsing is an unbounded pretraining signal: every known molecule can be annotated without experiments, so the recipe should scale well beyond the 250K molecules tested here.
  • A decisive control is overlap: the paper's claim of no generation-specific data still permits the same molecules to appear in pretraining, so a deduplication analysis would sharpen the transfer story.
  • The same deterministic-task template could extend to other molecular string formats such as SELFIES, or to stereochemical and 3D tasks that the paper explicitly leaves out.
  • Parsing pretraining is complementary to instruction tuning and could be combined with it rather than used as an alternative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces CLEANMOL, a two-stage framework that pretrains open-weight LLMs on five deterministic SMILES parsing tasks (functional group matching, ring counting, carbon chain length measurement, SMILES canonicalization, and fragment assembly) using 250K examples subsampled from ZINC250k with task-specific difficulty-based pruning and curriculum ordering. The pretrained model is then LoRA fine-tuned on three Mol-Instructions downstream tasks: retrosynthesis, reagent prediction, and forward reaction prediction. The paper reports that CLEANMOL improves parsing accuracy over strong general-purpose LLMs and yields consistent or best performance on downstream generation tasks compared with vanilla fine-tuning, a data-size-matched Mol-Instructions (SFT) baseline, and prior specialized models.

Significance. If the reported transfer is genuine, CLEANMOL offers a low-cost, scalable way to inject structural understanding into molecular LLMs, complementing existing instruction tuning and domain adaptation approaches. The paper contributes a well-motivated parsing task suite, a data-construction pipeline, and a fair data-size-controlled baseline (Mol-Instructions SFT), which strengthens the comparison. However, the central transfer claim hinges on the absence of pretraining/downstream overlap and on the reliability of the reported single-run numbers, both of which require verification before the result can be accepted.

major comments (4)
  1. [Section 5.1 / Section 4.1] The claim that CLEANMOL improves downstream generation through transfer of structural understanding is undermined by the absence of any overlap analysis between the ZINC250k molecules used for SMILES parsing pretraining (Section 4.1: 'The molecules are subsampled from the ZINC250k training dataset') and the Mol-Instructions molecules used for downstream evaluation (Section 5.1: 'We follow the data splits provided in Mol-Instructions'). Because the parsing suite includes canonicalization and fragment assembly, both of which train the model to emit exact canonical SMILES strings, any Mol-Instructions test molecule that also appears in the ZINC250k pretraining set could be memorized rather than understood. Please compute molecule-level overlap (e.g., canonical SMILES or InChI keys) between the CLEANMOL pretraining set and each Mol-Instructions test split, report the overlap fraction, and re-run the downstream evaluation on the disjoint subset. Without this, the abstract's causal interpretation ('enhance molecular generation capabilities') is not established.
  2. [Appendix B] All experimental results come from a single run, as stated in Appendix B ('our experiments are based on a single run'). Table 4 reports no error bars or significance tests, yet the paper claims CLEANMOL 'consistently improves performance across all backbones.' Several differences are small: for forward reaction prediction with LLaMA3.1-8B, +CLEANMOL is below Mol-Instructions (SFT) on RDK FTS (0.966 vs 0.967) and Levenshtein distance (1.37 vs 1.33), and on retrosynthesis the Exact match gap between +CLEANMOL and +Mol-Instructions (SFT) is 0.581 vs 0.541, within a range that could be noise. Please report means and standard deviations over at least 3 random seeds and, where appropriate, a paired significance test to support the consistency claim.
  3. [Section 4.2 / Table 3] The 'middle' pruning strategy is selected by comparing pruning variants on the fixed parsing test set (10K molecules from the ZINC250k test split, fixed across all experiments). Because the same test set is then used to report the final parsing accuracies in Tables 2, the reported numbers are subject to selection bias: the chose of the pruning method is itself optimized on the test set. Please hold out a separate validation split for method selection and report parsing performance on an untouched test split, or use nested cross-validation, so that the parsing results in Tables 2 and 3 reflect unbiased generalization.
  4. [Appendix A.2 / Fragment assembly] The fragment assembly task is only described by an instruction template ('Connect the following two SMILES fragments into a unified structure at their reactive sites'); the paper does not specify how the two fragments are generated, how the reactive sites are chosen, or whether the target product is unique given the fragments. If multiple valid assemblies are possible, the accuracy metric is not well-defined and the characterization of the task as 'deterministic' is incomplete. Please state the generation rule (e.g., cleavage of a single specific acyclic bond) and the evaluation criterion (exact match to a single target SMILES, or tolerance for alternate valid assemblies).
minor comments (6)
  1. [Section 1, Contributions] The word 'incoporating' in the third contribution bullet should be corrected to 'incorporating'.
  2. [Section 6, Related Work] The sentence 'most strategies introduce unambiguous supervision signals due to the non-determinism of molecular representations' appears to contain a wording error: non-determinism would cause ambiguous, not unambiguous, supervision signals. Please revise for clarity.
  3. [Section 3.2] It is unclear how the per-task mid-difficulty selections are combined into a single pretraining corpus: is a molecule kept if it is mid-difficulty for all five tasks, for any one task, or per task with a union of molecules? Please clarify the exact data construction rule.
  4. [Appendix C.1] The baseline name 'LlasMol' in Table 7 and the surrounding text appears to be a typo for 'LlaSMol' (Yu et al., 2024).
  5. [Throughout] The framework and dataset are sometimes written as 'CLEAN MOL' and sometimes as 'CLEANMOL'; please use one consistent spelling throughout the manuscript.
  6. [Section 3.2 / Experiments] The paper lists curriculum learning as a contribution, but Section 4.2 and Table 3 only ablate data pruning; there is no experiment comparing curriculum ordering versus random ordering. Please either add such an ablation or explicitly state that only the pruning component is empirically validated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the downstream Mol-Instructions evaluation is external to the CLEANMOL pretraining objective, and the only self-referential element is the held-out parsing evaluation, which is standard supervised evaluation rather than a prediction forced by construction.

full rationale

The paper's central transfer claim is that pretraining on deterministic SMILES parsing tasks improves downstream molecular generation (Section 5, Table 4). That claim is evaluated on Mol-Instructions data and metrics (Exact match, BLEU, MACCS/RDK/Morgan similarities, validity) that are not used to construct the parsing tasks or to fit any parameter of the pretraining objective; the downstream improvements are therefore an external, empirically contingent result rather than a quantity defined by the method's inputs. The parsing evaluation in Section 4.1 does reuse the same five task families employed in pretraining, so it measures whether the trained model can perform the trained skill on a held-out split; this is standard train/test evaluation, not a derivation in which the output equals the input by construction. The paper's citation of the authors' own Jang et al. (2024) is confined to motivation and explicit differentiation (footnote 1) and is not used as evidence or as a uniqueness constraint; no load-bearing self-citation, imported uniqueness theorem, or ansatz-via-citation pattern appears. The absence of a molecule-level deduplication check between ZINC250k pretraining molecules and Mol-Instructions test molecules is a genuine contamination risk for the causal 'transfer of structural understanding' interpretation, but it is a benchmark-validity concern, not a circularity of the derivation chain. Overall, the central derivation is self-contained against external benchmarks and the score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three unpaid premises: RDKit's automatic annotations are correct; structural knowledge learned on ZINC250k transfers to Mol-Instructions; and the hand-designed difficulty scores meaningfully rank parsing difficulty. The paper provides no invented physical entities. The free parameters are the unspecified pruning thresholds and difficulty definitions that determine the pretraining corpus.

free parameters (2)
  • mid-difficulty pruning selection
    The paper selects molecules with mid-level difficulty but does not specify the percentile range or how difficulty scores are aggregated across tasks (Section 3.2, Table 3).
  • task-specific difficulty measures
    Heuristics such as number of functional groups, number of rings, number of branches, and SMILES length are introduced as difficulty proxies without validation (Table 1).
assumptions (3)
  • domain assumption RDKit annotations (ring counts, functional group membership, canonical SMILES) are correct for all training examples.
    The dataset and parsing ground truth are generated entirely with RDKit (Section 3.1); errors in RDKit parsing would propagate to supervision.
  • domain assumption ZINC250k molecules are representative of the drug-like chemical space needed for Mol-Instructions downstream transfer.
    Pretraining uses ZINC250k while downstream uses Mol-Instructions; no distribution shift analysis is provided (Sections 4.1 and 5.1).
  • ad hoc to paper Single-epoch LoRA fine-tuning on 250K examples is sufficient to instill transferable structural understanding.
    Training is limited to one epoch with LoRA (Appendix B.1); the sufficiency of this choice is not justified beyond the reported results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Chemical Understanding of LLMs via SMILES Parsing." pith.science (2026). https://pith.science/paper/K7YLIDBM

@misc{pith2026250516340,
  author       = {Pith},
  title        = {Pith review of: Improving Chemical Understanding of LLMs via SMILES Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7YLIDBM}},
  note         = {Machine review of arXiv:2505.16340}
}
read the original abstract

Large language models (LLMs) are increasingly recognized as powerful tools for scientific discovery, particularly in molecular science. A fundamental requirement for these models is the ability to accurately understand molecular structures, commonly encoded in the SMILES representation. However, current LLMs struggle to interpret SMILES, even failing to carry out basic tasks such as counting molecular rings. To address this limitation, we introduce CLEANMOL, a novel framework that formulates SMILES parsing into a suite of clean and deterministic tasks explicitly designed to promote graph-level molecular comprehension. These tasks span from subgraph matching to global graph matching, providing structured supervision aligned with molecular structural properties. We construct a molecular pretraining dataset with adaptive difficulty scoring and pre-train open-source LLMs on these tasks. Our results show that CLEANMOL not only enhances structural comprehension but also achieves the best or competes with the baseline on the Mol-Instructions benchmark.

Figures

Figures reproduced from arXiv: 2505.16340 by the authors.

Figure 1
Figure 1. Overview of SMILES parsing. (a) Each column visualizes one of the five SMILES parsing tasks: functional group matching, ring counting, carbon chain length measurement, SMILES canonicalization, and fragment assembly. The highlighted tokens in the SMILES correspond to the substructures involved in each task. (b) Recent LLMs fail for SMILES parsing while the model trained with our CLEANMOL shows improvement. from molec… view at source ↗
Figure 2
Figure 2. Complex cases in SMILES parsing. The top green panels represent relatively simple cases, while the bottom red panels illustrate more complex examples with non-continuous substructures in SMILES. Orange and teal highlights correspond to tasks involving ring counting and functional group matching, respectively. 2 SMILES parsing task In this section, we introduce five SMILES parsing tasks designed to enhance the mappin… view at source ↗
Figure 3
Figure 3. Examples of CLEANMOL dataset. 😏 :Too easy :Too hard Canonical measure Pretrain Initial molecular dataset … Data pruning Ring measure 🤯 62🤯 47 31 44 53 68🤯 😏 3 2 5 4 4 4 🤯 😏 😏 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Overview of molecular data pruning and ranking. Each number represents the task-specific diffi￾culty score assigned to a molecule, as defined in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Data scale analysis for SMILES parsing. the accuracy of the SMILES parsing task for 10K, 20K, and 50K data settings per task in the same setting in Section 4.1. We provide the results in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Data scale analysis for retrosynthesis. hancing molecular language modeling. These im￾provements suggest that pre-training on clean and deterministic CLEANMOL dataset facilitates the model’s structural understanding required for gen￾eration tasks. Notably, integrating …
Figure 8
Figure 8. Figure 8: Ring counting Answer only with the corresponding integer number without any other information. **Question:** Assess the SMILES below and report how many rings consist of [RING SIZE] atoms. Give me the integer only. **SMILES:** [SMILES] **SIZE OF RINGS:** [RING SIZE] **…
Figure 9
Figure 9. Figure 9: An instruction format of chain length mea￾surement. A.2 Global graph matching This category includes tasks that operate on a global level: SMILES canonicalization and frag￾ment assembly. Unlike subgraph matching, these tasks require full-graph interpretation, where suc…
Figure 11
Figure 11. Figure 11: An instruction format of SMILES assem￾bly. (Daniel Han and team, 2023). Additionally, we used the packages including rouge-score==0.1.2 and nltk==3.8.1. B.1 SMILES parsing Here, we describe the detailed settings for the SMILES parsing experiments in Section 4, includ￾…
Figure 10
Figure 10. Figure 10: An instruction format of SMILES canoni￾calization. Fragment assembly. Fragment assembly eval￾uates whether the model can reconstruct a full molecule from two disconnected SMILES frag￾ments. This task tests global molecular coherence and the model’s ability to resolve …

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Reaction-Network-Level Discovery of Ammonia Synthesis Catalysts via Ten-Million-Scale Generative Exploration

    physics.chem-ph 2026-06 unverdicted novelty 7.0 of 10

    Ten-million-scale generative Transformers with ML potentials map compatibility across N*, NH*, NNH*, and HNNH* to discover 279 ammonia synthesis catalyst candidates, recovering Fe/Ru motifs and identifying new familie...

  2. Exploring Modularity of Agentic Systems for Drug Discovery

    cs.LG 2025-06 conditional novelty 4.0 of 10

    On 26 chemistry questions, swapping the LLM, agent type, or prompt in an LLM agent changes its scores so much that the system cannot be treated as modular.

Reference graph

Works this paper leans on

59 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    Zachary Ankner, Cody Blakeney, Kartik Sreenivasan, Max Marion, Matthew L Leavitt, and Mansheej Paul. 2024. Perplexed by perplexity: Perplexity-based data pruning with small reference models. arXiv preprint arXiv:2405.20541

    arXiv 2024

  2. [2]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41--48

    2009

  3. [3]

    He Cao, Zijing Liu, Xingyu Lu, Yuan Yao, and Yu Li. 2023. https://arxiv.org/abs/2311.16208 Instructmol: Multi-modal integration for building a versatile and reliable molecular assistant in drug discovery . Preprint, arXiv:2311.16208

    arXiv 2023

  4. [4]

    Cayque Monteiro Castro Nascimento and Andr \'e Silva Pimentel. 2023. https://doi.org/10.1021/acs.jcim.3c00285 Do large language models understand chemistry? a conversation with chatgpt . Journal of Chemical Information and Modeling, 63(6):1649--1655

    doi:10.1021/acs.jcim.3c00285 2023

  5. [5]

    Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, and 1 others. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ See https://vicuna. lmsys. org (accessed 14 April 2023), 2(3):6

    2023

  6. [6]

    Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023 a . https://proceedings.mlr.press/v202/christofidellis23a.html Unifying molecular and textual representations via multi-task language modelling . In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of...

    work page 2023

  7. [7]

    Dimitrios Christofidellis, Giorgio Giannone, Jannis Born, Ole Winther, Teodoro Laino, and Matteo Manica. 2023 b . https://proceedings.mlr.press/v202/christofidellis23a.html Unifying molecular and textual representations via multi-task language modelling . In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of...

    work page 2023

  8. [8]

    Michael Han Daniel Han and Unsloth team. 2023. http://github.com/unslothai/unsloth Unsloth

    2023

Show all 59 references
  1. [9]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248--255. Ieee

    2009

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human...

    2019

  3. [11]

    Joseph L Durant, Burton A Leland, Douglas R Henry, and James G Nourse. 2002. Reoptimization of mdl keys for use in drug discovery. Journal of chemical information and computer sciences, 42(6):1273--1280

    2002

  4. [12]

    Carl Edwards, Tuan Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.26 Translation between molecules and natural language . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pa...

    2022 doi

  5. [13]

    Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Huajun Chen. 2024. https://openreview.net/forum?id=Tlsdsb6l9n Mol-instructions: A large-scale biomolecular instruction dataset for large language models . In The Twelfth International ...

    2024

  6. [14]

    Veronika Ganeeva, Andrey Sakhovskiy, Kuzma Khrabrov, Andrey Savchenko, Artur Kadurin, and Elena Tutubalina. 2024. Lost in translation: Chemical language models and the misunderstanding of molecule structures. In Findings of the Association for Computational Linguistics: EMNLP ...

    2024

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  8. [16]

    Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, and Benjamin Bossan. 2022. Accelerate: Training and inference at scale made simple, efficient and adaptable. https://github.com/huggingface/accelerate

    2022

  9. [17]

    Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio C \'e sar Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, and 1 others. 2023. Textbooks are all you need. arXiv preprint arXiv:2306.11644

    2023 arXiv

  10. [18]

    Taicheng Guo, Bozhao Nan, Zhenwen Liang, Zhichun Guo, Nitesh Chawla, Olaf Wiest, Xiangliang Zhang, and 1 others. 2023. What can large language models do in chemistry? a comprehensive benchmark on eight tasks. Advances in Neural Information Processing Systems, 36:59662--59688

    2023

  11. [19]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  12. [20]

    John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. 2012. Zinc: a free tool to discover chemistry for biology. Journal of chemical information and modeling, 52(7):1757--1768

    2012

  13. [21]

    Yunhui Jang, Jaehyung Kim, and Sungsoo Ahn. 2024. Chain-of-thoughts for molecular understanding. arXiv preprint arXiv:2410.05610

    2024 arXiv

  14. [22]

    Xiaohong Ji, Zhen Wang, Zhifeng Gao, Hang Zheng, Linfeng Zhang, Guolin Ke, and Weinan E. 2024. https://openreview.net/forum?id=64V40K2fDv Exploring molecular pretraining model at scale . In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  15. [23]

    Mario Krenn, Florian H \"a se, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. 2020. Self-referencing embedded strings (selfies): A 100\ Machine Learning: Science and Technology, 1(4):045024

    2020

  16. [24]

    Greg Landrum, Paolo Tosco, Brian Kelley, Ricardo Rodriguez, David Cosgrove, Riccardo Vianello, sriniker, Peter Gedeck, Gareth Jones, NadineSchneider, Eisuke Kawashima, Dan Nealschneider, Andrew Dalke, Matt Swain, Brian Cole, Samo Turk, Aleksandr Savelev, Alain Vaucher, Maciej ...

    2024 doi

  17. [25]

    Han Li, Ruotian Zhang, Yaosen Min, Dacheng Ma, Dan Zhao, and Jianyang Zeng. 2023. A knowledge-guided pre-training framework for improving molecular representation learning. Nature Communications, 14(1):7568

    2023

  18. [26]

    Jiatong Li, Yunqing Liu, Wenqi Fan, Xiao-Yong Wei, Hui Liu, Jiliang Tang, and Qing Li. 2024. https://doi.org/10.1109/tkde.2024.3393356 Empowering molecule discovery for molecule-caption translation with large language models: A chatgpt perspective . IEEE Transactions on Knowle...

    2024

  19. [27]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  20. [28]

    Zequn Liu, Wei Zhang, Yingce Xia, Lijun Wu, Shufang Xie, Tao Qin, Ming Zhang, and Tie-Yan Liu. 2023 a . https://doi.org/10.18653/v1/2023.acl-short.138 M ol XPT : Wrapping molecules with text for generative pre-training . In Proceedings of the 61st Annual Meeting of the Associa...

    2023 doi

  21. [29]

    Zhiyuan Liu, Sihang Li, Yanchen Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.966 M ol CA : Molecular graph-language modeling with cross-modal projector and uni-modal adapter . In Proceedings of th...

    2023 doi

  22. [30]

    Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, and 1 others. 2024. Starcoder 2 and the stack v2: The next generation. arXiv preprint arXiv:2402.19173

    2024 arXiv

  23. [31]

    Jieyu Lu and Yingkai Zhang. 2022. Unified deep learning model for multitask reaction predictions with explanation. Journal of chemical information and modeling, 62(6):1376--1387

    2022

  24. [32]

    Max Marion, Ahmet \"U st \"u n, Luiza Pozzobon, Alex Wang, Marzieh Fadaee, and Sara Hooker. 2023. When less is more: Investigating data pruning for pretraining llms at scale. arXiv preprint arXiv:2309.04564

    2023 arXiv

  25. [33]

    Frederic P Miller, Agnes F Vandome, and John McBrewster. 2009. Levenshtein distance: Information theory, computer science, string (computer science), string metric, damerau? levenshtein distance, spell checker, hamming distance

    2009

  26. [34]

    OpenAI and Josh Achiam et al. 2024. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2024 arXiv

  27. [35]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

    2002

  28. [36]

    Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.70 B io T 5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations . In Proceedings of...

    2023 doi

  29. [37]

    Kristina Preuer, Philipp Renz, Thomas Unterthiner, Sepp Hochreiter, and G \"u nter Klambauer. 2018. https://doi.org/10.1021/acs.jcim.8b00234 Fr \'e chet chemnet distance: A metric for generative models for molecules in drug discovery . Journal of Chemical Information and Model...

    2018 doi

  30. [38]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020 a . Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1--67

    2020

  31. [39]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020 b . http://jmlr.org/papers/v21/20-074.html Exploring the limits of transfer learning with a unified text-to-text transformer . Journal of Machine ...

    2020

  32. [40]

    David Rogers and Mathew Hahn. 2010. Extended-connectivity fingerprints. Journal of chemical information and modeling, 50(5):742--754

    2010

  33. [41]

    Nadine Schneider, Roger A Sayle, and Gregory A Landrum. 2015. Get your atoms in order - an open-source implementation of a novel and robust molecular canonicalization algorithm. Journal of chemical information and modeling, 55(10):2111--2120

    2015

  34. [42]

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A large language model for science. arXiv preprint arXiv:2211.09085

    2022 arXiv

  35. [43]

    Kushal Tirumala, Daniel Simig, Armen Aghajanyan, and Ari Morcos. 2023. D4: Improving llm pretraining via document de-duplication and diversification. Advances in Neural Information Processing Systems, 36:53983--53995

    2023

  36. [44]

    Tloen. 2023. Alpaca-lora. https://github.com/tloen/alpaca-lora

    2023

  37. [45]

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. 2020. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl

    2020

  38. [46]

    Jin-Mao Wei, Xiao-Jie Yuan, Qing-Hua Hu, and Shu-Qin Wang. 2010. A novel measure for evaluating classifiers. Expert Systems with Applications, 37(5):3799--3809

    2010

  39. [47]

    David Weininger. 1988. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of Chemical Information and Computer Sciences, 28(1):31--36

    1988

  40. [48]

    David Weininger, Arthur Weininger, and Joseph L Weininger. 1989. Smiles. 2. algorithm for generation of unique smiles notation. Journal of chemical information and computer sciences, 29(2):97--101

    1989

  41. [49]

    Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzm \'a n, Armand Joulin, and Edouard Grave. 2020. https://aclanthology.org/2020.lrec-1.494/ CCN et: Extracting high quality monolingual datasets from web crawl data . In Proceedings of the Twe...

    2020

  42. [50]

    White, Glen M

    Andrew D. White, Glen M. Hocky, Heta A. Gandhi, Mehrad Ansari, Sam Cox, Geemi P. Wellawatte, Subarna Sasmal, Ziyue Yang, Kangxin Liu, Yuvraj Singh, and Willmor J. Peña Ccoa. 2023. https://doi.org/10.1039/D2DD00087C Assessment of chemistry knowledge in large language models tha...

    2023 doi

  43. [51]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and ...

    2020

  44. [52]

    Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. 2018. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9(2):513--530

    2018

  45. [53]

    Yingce Xia, Peiran Jin, Shufang Xie, Liang He, Chuan Cao, Renqian Luo, Guoqing Liu, Yue Wang, Zequn Liu, Yuan-Jyue Chen, Zekun Guo, Yeqi Bai, Pan Deng, Yaosen Min, Ziheng Lu, Hongxia Hao, Han Yang, Jielan Li, Chang Liu, and 27 others. 2025. https://arxiv.org/abs/2502.07527 Nat...

    2025 arXiv

  46. [54]

    Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.385 Baize: An open-source chat model with parameter-efficient tuning on self-chat data . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processi...

    2023 doi

  47. [55]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  48. [56]

    Baker, Ziqi Chen, Xia Ning, and Huan Sun

    Botao Yu, Frazier N. Baker, Ziqi Chen, Xia Ning, and Huan Sun. 2024. https://openreview.net/forum?id=lY6XTF9tPv Lla SM ol: Advancing large language models for chemistry with a large-scale, comprehensive, high-quality instruction tuning dataset . In First Conference on Language...

    2024

  49. [57]

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. 2023. https://openreview.net/forum?id=6K2RM6wVqKu Uni-mol: A universal 3d molecular representation learning framework . In The Eleventh International Conference on Learni...

    2023

  50. [58]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  51. [59]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.