---
license: cc-by-4.0
pretty_name: Olfactory Perception (OP) Benchmark
language:
- en
- ar
- bn
- de
- el
- es
- fa
- fr
- hi
- it
- ja
- ko
- pl
- pt
- ru
- sv
- sw
- tr
- uk
- zh
task_categories:
- question-answering
tags:
- olfaction
- smell
- chemistry
- SMILES
- benchmark
- llm-evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: Benchmark/OP_Benchmark.csv
dataset_info:
  config_name: default
  features:
  - name: question_ID
    dtype: string
  - name: compound.name_1
    dtype: string
  - name: SMILES_1
    dtype: string
  - name: compound.name_2
    dtype: string
  - name: SMILES_2
    dtype: string
  - name: OPTIONS
    dtype: string
  - name: question_category
    dtype: string
  - name: prompt.1
    dtype: string
  - name: prompt.2
    dtype: string
  - name: answer
    dtype: string
  - name: other_info
    dtype: string
---

# Olfactory Perception (OP) Benchmark

The OP benchmark evaluates what large language models know about smell. It contains **1,010 questions** across eight task categories, each asked in two formats: with **compound names** and with **isomeric SMILES**.

| Task | Code | Questions | Format |
|---|---|---|---|
| Odor classification | OC | 175 | Odorous / odorless |
| Odor primary descriptor | OPD | 175 | 1 of 4 descriptors |
| Odor intensity | OIn | 175 | Which of two compounds is more intense (+ 0–100 ratings) |
| Odor pleasantness | OPl | 175 | Which of two compounds is more pleasant (+ 0–100 ratings) |
| Rate-all-that-apply | RATA | 100 | All applicable descriptors from 138 |
| Odor similarity of mixtures | OS | 100 | 4-level similarity (+ 0–1 distance) |
| Olfactory receptor activation | ORA | 80 | All activated receptors from 4–10 candidates |
| Smell identification test | SIT | 30 | 1 of 4 odor sources |

The release also includes the RATA task translated into 20 further languages (`Benchmark/4-Intermediate_1/Translations_GPT5.2/`).

- **Paper:** *Benchmark for Assessing Olfactory Perception of Large Language Models*, NeurIPS 2026 (Evaluations and Datasets Track)
- **Code, prompts, evaluation and analysis scripts:** https://github.com/Satarifard/Olfactory-Perception-Benchmark
- **Authors:** Eftychia Makri (Yale University), Nikolaos Nakis (Yale University), Laura Sisson (Patina), Geetanjali Minsky (MIT Media Lab), Leandros Tassiulas (Yale University), Vahid Satarifard (Yale University), Nicholas A. Christakis (Yale University)
- **Contact:** eftychia.makri@yale.edu, vahid.satarifard@yale.edu

## Columns

| Column | Description |
|---|---|
| `question_ID` | Unique question identifier (UUID) |
| `question_category` | `odor_classification`, `primary_odor_descriptor`, `odor_intensity`, `odor_pleasantness`, `rata`, `mixture_similarity`, `or_activation`, or `smell_identification` |
| `compound.name_1`, `SMILES_1` | First compound, or first mixture (semicolon-separated) |
| `compound.name_2`, `SMILES_2` | Second compound or mixture (paired and mixture tasks only) |
| `OPTIONS` | Answer options (semicolon-separated) |
| `prompt.1` | Full prompt with isomeric SMILES |
| `prompt.2` | Full prompt with compound names |
| `answer` | Ground-truth answer (semicolon-separated for RATA and ORA) |
| `other_info` | Supplementary values, e.g. human ratings for OIn/OPl |


## Sources

Questions are built from: Mayhew et al. 2022 (OC); the IFRA Fragrance Ingredient Glossary, 2020 (OPD); Keller et al. 2017 (OIn, OPl); GS-LF as used by Lee et al. 2023 (RATA); Snitz et al. 2013, Bushdid et al. 2014, and Ravia et al. 2020 (OS); M2OR, Lalis et al. 2024 (ORA); and the Leibniz-LSB@TUM Odorant Database, Kreissl et al. 2022 (SIT). Please cite the original sources as well when using the corresponding tasks.

## License

CC BY 4.0.

## Citation

```bibtex
@inproceedings{makri2026op,
  title     = {Benchmark for Assessing Olfactory Perception of Large Language Models},
  author    = {Makri, Eftychia and Nakis, Nikolaos and Sisson, Laura and Minsky, Geetanjali and Tassiulas, Leandros and Satarifard, Vahid and Christakis, Nicholas A.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}
```