Datasets:
Update dataset card with paper link, code link, and metadata
#1
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,21 +1,23 @@
|
|
| 1 |
---
|
| 2 |
-
license:
|
| 3 |
-
tags:
|
| 4 |
-
- chemistry
|
| 5 |
-
- supramolecular
|
| 6 |
-
- host-guest
|
| 7 |
-
- molecular-recognition
|
| 8 |
-
- macrocyclic
|
| 9 |
-
- continued-pretraining
|
| 10 |
size_categories:
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
configs:
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
---
|
| 20 |
|
| 21 |
# SupraBench
|
|
@@ -25,6 +27,9 @@ configs:
|
|
| 25 |
tasks plus an auxiliary vision task, and ships a domain text corpus for
|
| 26 |
domain-adaptive pretraining (DAPT).
|
| 27 |
|
|
|
|
|
|
|
|
|
|
| 28 |
> Supramolecular chemistry studies non-covalent host-guest assemblies that
|
| 29 |
> underpin drug delivery, chemical sensing, and in-vivo toxin sequestration.
|
| 30 |
> Designing host-guest systems is slow (days of dry-lab verification per pair);
|
|
@@ -45,6 +50,17 @@ Each task dataset has a `test` split (merged records) and a `cb7` split (the
|
|
| 45 |
CB[7] supplement, for add-on evaluation). Each `base`/`fewshot`/`cot` rendering is
|
| 46 |
tagged by the `prompt_strategy` field.
|
| 47 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
## Dataset statistics
|
| 49 |
|
| 50 |
| Task | # Samples |
|
|
@@ -79,32 +95,6 @@ column. Arrows give the optimization direction.
|
|
| 79 |
| Gemini-3-Flash | **1.248** | **1.679** | **0.498** | **0.647** | **0.350** | **0.470** | **0.506** | 0.067 | 0.118 |
|
| 80 |
| DeepSeek-v4 | *1.433* | *1.994* | *0.461* | *0.730* | *0.309* | *0.381* | *0.500* | *0.090* | *0.141* |
|
| 81 |
|
| 82 |
-
### Few-Shot
|
| 83 |
-
|
| 84 |
-
| Model | BAP MAE down | BAP RMSE down | TBS ACC up | TBS Regret down | SID F1 up | SID B.Acc up | HGD Recall up | HGD Prec up | HGD F1 up |
|
| 85 |
-
|---|---|---|---|---|---|---|---|---|---|
|
| 86 |
-
| Qwen3.5-9B | 3.650 | 4.820 | 0.370 | 0.951 | 0.154 | 0.150 | 0.000 | 0.022 | 0.042 |
|
| 87 |
-
| Qwen3.5-27B | 2.258 | 3.256 | 0.392 | 0.889 | 0.178 | 0.257 | 0.636 | **0.585** | **0.580** |
|
| 88 |
-
| Llama3.1-8B | 5.504 | 6.940 | 0.283 | 1.227 | 0.142 | 0.182 | 0.655 | 0.369 | 0.456 |
|
| 89 |
-
| Llama3.1-70B | 1.774 | 2.359 | 0.354 | 1.026 | 0.144 | 0.185 | 0.631 | *0.474* | *0.531* |
|
| 90 |
-
| GPT-5.4-Mini | 1.958 | 2.808 | 0.430 | 0.824 | 0.141 | *0.291* | 0.542 | 0.228 | 0.307 |
|
| 91 |
-
| GPT-5.4-Nano | 2.176 | 2.894 | 0.419 | 0.819 | 0.190 | 0.270 | 0.532 | 0.095 | 0.152 |
|
| 92 |
-
| Gemini-3-Flash | **1.257** | **1.702** | **0.513** | **0.619** | **0.389** | **0.421** | *0.660* | 0.364 | 0.448 |
|
| 93 |
-
| DeepSeek-v4 | *1.618* | *2.276* | *0.470* | *0.713* | *0.203* | 0.225 | **0.720** | 0.303 | 0.352 |
|
| 94 |
-
|
| 95 |
-
### CoT
|
| 96 |
-
|
| 97 |
-
| Model | BAP MAE down | BAP RMSE down | TBS ACC up | TBS Regret down | SID F1 up | SID B.Acc up | HGD Recall up | HGD Prec up | HGD F1 up |
|
| 98 |
-
|---|---|---|---|---|---|---|---|---|---|
|
| 99 |
-
| Qwen3.5-9B | 3.664 | 4.885 | 0.382 | 0.944 | 0.167 | 0.197 | 0.300 | 0.039 | 0.068 |
|
| 100 |
-
| Qwen3.5-27B | 2.438 | 3.468 | 0.398 | 0.898 | 0.254 | *0.415* | **0.526** | 0.051 | 0.092 |
|
| 101 |
-
| Llama3.1-8B | 4.911 | 6.279 | 0.293 | 1.220 | 0.154 | 0.153 | 0.380 | **0.102** | **0.144** |
|
| 102 |
-
| Llama3.1-70B | 1.833 | 2.512 | 0.373 | 0.985 | 0.106 | 0.380 | 0.421 | 0.055 | 0.097 |
|
| 103 |
-
| GPT-5.4-Mini | 2.036 | 2.887 | 0.429 | 0.828 | 0.220 | 0.282 | 0.444 | 0.080 | 0.129 |
|
| 104 |
-
| GPT-5.4-Nano | 2.160 | 2.881 | 0.410 | 0.822 | 0.174 | 0.257 | 0.492 | 0.056 | 0.098 |
|
| 105 |
-
| Gemini-3-Flash | **1.261** | **1.723** | **0.510** | **0.609** | **0.331** | **0.432** | 0.512 | 0.062 | 0.110 |
|
| 106 |
-
| DeepSeek-v4 | *1.541* | *2.183* | *0.445* | *0.743* | *0.307* | 0.414 | *0.522* | *0.080* | *0.134* |
|
| 107 |
-
|
| 108 |
Takeaways: frontier proprietary LLMs (Gemini-3-Flash, DeepSeek-v4) lead the
|
| 109 |
quantitative tasks, yet every task leaves substantial headroom; no single
|
| 110 |
prompting strategy is universally best; and CoT amplifies rather than fixes the
|
|
@@ -119,3 +109,14 @@ articles subject to each article's individual license; molecular structures use
|
|
| 119 |
[OPSIN](https://github.com/dan2097/opsin).
|
| 120 |
|
| 121 |
If you use SupraBench, please cite the paper and the upstream data sources.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
size_categories:
|
| 4 |
+
- 100K<n<1M
|
| 5 |
+
task_categories:
|
| 6 |
+
- text-generation
|
| 7 |
+
tags:
|
| 8 |
+
- chemistry
|
| 9 |
+
- supramolecular
|
| 10 |
+
- host-guest
|
| 11 |
+
- molecular-recognition
|
| 12 |
+
- macrocyclic
|
| 13 |
+
- continued-pretraining
|
| 14 |
configs:
|
| 15 |
+
- config_name: default
|
| 16 |
+
data_files:
|
| 17 |
+
- split: raw
|
| 18 |
+
path: data/raw-*.parquet
|
| 19 |
+
- split: filtered
|
| 20 |
+
path: data/filtered-*.parquet
|
| 21 |
---
|
| 22 |
|
| 23 |
# SupraBench
|
|
|
|
| 27 |
tasks plus an auxiliary vision task, and ships a domain text corpus for
|
| 28 |
domain-adaptive pretraining (DAPT).
|
| 29 |
|
| 30 |
+
- 📄 **Paper:** [SupraBench: A Benchmark for Supramolecular Chemistry](https://huggingface.co/papers/2606.13477)
|
| 31 |
+
- 💻 **Code:** [Tianyi-Billy-Ma/SupraBench](https://github.com/Tianyi-Billy-Ma/SupraBench)
|
| 32 |
+
|
| 33 |
> Supramolecular chemistry studies non-covalent host-guest assemblies that
|
| 34 |
> underpin drug delivery, chemical sensing, and in-vivo toxin sequestration.
|
| 35 |
> Designing host-guest systems is slow (days of dry-lab verification per pair);
|
|
|
|
| 50 |
CB[7] supplement, for add-on evaluation). Each `base`/`fewshot`/`cot` rendering is
|
| 51 |
tagged by the `prompt_strategy` field.
|
| 52 |
|
| 53 |
+
## Sample usage
|
| 54 |
+
|
| 55 |
+
To run a task against a model using the provided source code:
|
| 56 |
+
|
| 57 |
+
```bash
|
| 58 |
+
uv run python src/main.py \
|
| 59 |
+
--task-config configs/tasks/bap_base.yaml \
|
| 60 |
+
--model-config configs/models/openrouter_qwen35_27b.yaml \
|
| 61 |
+
--output-dir outputs/
|
| 62 |
+
```
|
| 63 |
+
|
| 64 |
## Dataset statistics
|
| 65 |
|
| 66 |
| Task | # Samples |
|
|
|
|
| 95 |
| Gemini-3-Flash | **1.248** | **1.679** | **0.498** | **0.647** | **0.350** | **0.470** | **0.506** | 0.067 | 0.118 |
|
| 96 |
| DeepSeek-v4 | *1.433* | *1.994* | *0.461* | *0.730* | *0.309* | *0.381* | *0.500* | *0.090* | *0.141* |
|
| 97 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
Takeaways: frontier proprietary LLMs (Gemini-3-Flash, DeepSeek-v4) lead the
|
| 99 |
quantitative tasks, yet every task leaves substantial headroom; no single
|
| 100 |
prompting strategy is universally best; and CoT amplifies rather than fixes the
|
|
|
|
| 109 |
[OPSIN](https://github.com/dan2097/opsin).
|
| 110 |
|
| 111 |
If you use SupraBench, please cite the paper and the upstream data sources.
|
| 112 |
+
|
| 113 |
+
## Citation
|
| 114 |
+
|
| 115 |
+
```bibtex
|
| 116 |
+
@article{ma2026suprabench,
|
| 117 |
+
title = {SupraBench: A Benchmark for Supramolecular Host--Guest Chemistry Reasoning in Large Language Models},
|
| 118 |
+
author = {Ma, Tianyi and Ma, Yijun and Wang, Zehong and Sun, Weixiang and Li, Ziming and Schmidt, Connor R. and Zhang, Chuxu and Webber, Matthew J. and Ye, Yanfang},
|
| 119 |
+
year = {2026},
|
| 120 |
+
note = {arXiv preprint, link coming soon}
|
| 121 |
+
}
|
| 122 |
+
```
|