Syzseisus/S3-Prompt

S³-Prompt — Paper Results & Checkpoints

Released artifacts for S³-Prompt: Resolving Activity Cliffs via Structure-Disentangled Molecular Graph-Text Alignment (CIKM '26).

This repository holds the runs behind the numbers reported in Tables 2, 3, 5 and 7 of the paper: configuration, raw predictions, a human-readable metric summary, and the trained checkpoint (GNN + projectors, plus LoRA adapters where applicable).

• 💻 Code: https://github.com/Syzseisus/S3-Prompt
• 📊 Preprocessed data: https://huggingface.co/datasets/Syzseisus/S3-Prompt

Layout

results_final/
├── {BACE,BBBP,ESOL,LIPOPHILICITY}/
│   └── {A..G}_{FP,Sem}/k{2,8}_{full,short}/
│       ├── args.json        # exact run configuration
│       ├── result*.json     # predictions + metrics
│       ├── metric.txt       # human-readable summary of every metric
│       └── epoch_*.pt       # checkpoint (when retained)
├── Table_3.md               # classification component ablation (AUC)
├── Table_5.md               # sensitivity to shots k / instruction level
└── Table_7.md               # regression component ablation (RMSE)

The folder name encodes the ablation index (A–G) and retrieval modality (FP topological / Sem semantic). The subfolder k{k}_{full,short} is the number of shots and the instruction level; the paper setting is k8_full, and the other variants belong to the sensitivity study (Table 5).

Ablation index

Index	Decomposition	LoRA	Triplet loss
A	none	—	—
B	text	—	—
C	text	✓	—
D	soft	—	—
E	soft	—	✓
F	soft	✓	—
G	soft	✓	✓

Coverage

27 runs, covering 31 of the 40 cells in Tables 2, 3, 5 and 7 (some cells share a run).

Table	Cells	Released	Missing
Table 2 — main results (classification)	4	2	BACE · S³-Prompt (Sem) 0.831; BBBP · S³-Prompt (FP) 0.903
Table 3 — classification ablation	14	14	—
Table 5 — sensitivity (k / instruction)	8	6	BACE (Index E, FP): k=2 short 0.845, k=2 full 0.873
Table 7 — regression ablation	14	9	LIPO Index D 0.9574, E 1.0056, F 1.0236; ESOL Index D 1.5855, F 1.3788

The two released Table 2 cells are the same runs as Table 3 (BACE Index E · FP, BBBP Index G · Sem). Table 6 (regression main results) is not separately covered: its ESOL · S³-Prompt (Sem) cell is the Table 7 ESOL Index G run; the other three cells have no released run.

Notes on the released runs

• Regression Index D and F (Table 7) are withheld. The runs behind these cells were trained with the triplet loss enabled (lambda_loss = 0.1), so they do not match the "no triplet" setting the table describes. They are therefore not released and are listed as missing above. The sweep script run_regression.sh trains D and F with lambda_loss = 0.
• BACE Index F (Table 3): the released run scores AUC 0.8770; the paper prints 0.878.
• Regression LoRA runs (Index C and G): trained with LoRA rank 8 / alpha 16. The paper (and the classification runs) use rank 12 / alpha 24.
• Checkpoints not released. Index A and B are inference-only and have no checkpoint. The following trained runs ship results without epoch_*.pt: BACE C (FP); BBBP C, D, E (Sem); BACE E k8_short (Table 5); BBBP G k2_short and k2_full (Table 5).

Usage

from huggingface_hub import snapshot_download
snapshot_download("Syzseisus/S3-Prompt", repo_type="model", local_dir="S3-Prompt")

Citation

@inproceedings{s3prompt2026,
  title     = {S$^3$-Prompt: Resolving Activity Cliffs via Structure-Disentangled Molecular Graph-Text Alignment},
  author    = {Cho, Wooseong Syz and Park, Ha-Myung},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM '26)},
  year      = {2026},
  address   = {Rome, Italy},
  publisher = {ACM},
  doi       = {10.1145/3799682.3841121},
}