opbenchmark commited on
Commit
a6c46cc
·
verified ·
1 Parent(s): 5316254

Update croissant_rai_Olfactory-Perception-Benchmark.json

Browse files
croissant_rai_Olfactory-Perception-Benchmark.json CHANGED
@@ -73,11 +73,11 @@
73
  "multilingual",
74
  "🇺🇸 Region: US"
75
  ],
76
- "rai:dataLimitations": "OP is an evaluation-only benchmark not intended for model training or fine-tuning. The benchmark uses standardized descriptor vocabularies and discrete response formats (multiple choice, multilabel selection from fixed lists) which enable reproducible automatic evaluation but do not capture the full richness of odor perception, including contextual effects, individual variation, free-form descriptions, or temporal/concentration dynamics. Ground-truth labels reflect agreement with established olfactory science datasets and should not be interpreted as evidence of mechanistic olfactory understanding. The RATA descriptor lexicon (138 GS-LF descriptors) and the M2OR receptor set are themselves products of specific scientific traditions and may not generalize to all olfactory perception phenomena. The 100 RATA molecules overlap entirely with the GS-LF training set used for the Principal Odor Map; benchmark scores against specialist models must therefore be interpreted with leakage in mind (a fair-retrained POM comparison is provided). The mixture-similarity task (OS) inherits ordinal-scale binning decisions from the source datasets. Multilingual RATA variants rely on automated GPT-5.2 translation; native-speaker validation was conducted for Italian, Persian, Greek, and Russian only — subtle connotation shifts may remain in the other 16 non-English languages. The benchmark is strictly not recommended for: clinical or safety-critical olfactory deployment (e.g., gas detection, food safety certification), training fragrance-design models without independent evaluation, evidence of mechanistic olfactory understanding, or commercial applications without further validation.",
77
- "rai:dataBiases": "OP inherits the biases of its source datasets. (i) Compound coverage: molecules are drawn from peer-reviewed olfactory science resources (GS-LF, M2OR, IFRA FIG, DREAM Challenge data, Leibniz-LSB@TUM odorant database) which over-represent fragrance and flavor industry interests and Western culinary traditions; less commercially relevant odorants are systematically underrepresented. (ii) Descriptor lexicon: the 138-descriptor RATA vocabulary follows GS-LF conventions developed primarily in English-speaking and European perfumery contexts, biasing semantic categories toward Western olfactory framings. (iii) Receptor coverage: the M2OR-derived ORA task covers human olfactory receptors that have been experimentally characterized; rare or recently discovered receptors are absent. Receptor-ligand pairs reflect the experimental priorities of academic labs and pharmaceutical screens. (iv) Real-world source bias: SIT items (30 foods and beverages) reflect items prevalent in European and North American food science literature. (v) Multilingual coverage: family-level coverage is unbalanced (more Indo-European than non-Indo-European languages); GPT-5.2 translation quality reflects the model's training distribution and is expected to be stronger for higher-resourced languages. Cross-cultural residual analysis identifies systematic over-performance on Romance and Germanic languages and under-performance on Korean, Arabic, Persian, Hindi, and Bengali, consistent with cross-linguistic differences in dedicated olfactory vocabulary.",
78
- "rai:personalSensitiveInformation": "The dataset contains no directly Personally Identifiable Information (PII) — no names, contact details, faces, biometric data, health records, or national identifiers. The dataset does not contain personal disclosures from human subjects. Ground-truth perceptual ratings (intensity, pleasantness, mixture similarity) are aggregated mean values from the source datasets (DREAM Challenge psychophysical studies; Snitz et al.; Bushdid et al.; Ravia et al.) rather than individual-level responses; individual-level data was not collected by the authors of this work. The multilingual translation process was performed by an automated language model (GPT-5.2) and did not involve human subjects research; native-speaker reviewers (Italian, Persian, Greek, Russian) participated in a sample-review capacity. No crowdsourcing was involved in benchmark construction.",
79
  "rai:dataUseCases": "OP is intended to measure factual olfactory knowledge in large language models — specifically, whether LLMs can reason from molecular information (compound names or isomeric SMILES) to perceptual properties of smell. Validated use cases include: (1) any-overlap accuracy on single-answer tasks (OC, OPD, OIn, OPl, OS, SIT) for closed-set olfactory question answering; (2) per-question multilabel F1 on RATA and ORA for descriptor and receptor knowledge evaluation; (3) Pearson correlation against human psychophysical ratings on intensity, pleasantness, and mixture similarity; (4) molecular-representation comparison via the dual-prompting (isomeric SMILES vs. compound name) design; (5) cross-lingual evaluation of olfactory descriptor knowledge via the 21-language RATA variant and ensemble approaches. Validity is established through benchmarking of 21 model configurations across 6 providers, paired statistical tests with Holm-Bonferroni correction, hierarchical bootstrap confidence intervals, and comparison to specialist models (POM ensemble, fair-retrained). Use cases NOT validated and explicitly out of scope include: training or fine-tuning olfactory prediction models (test-set integrity is required for benchmark validity); clinical or industrial olfactory deployment; fragrance or flavor product development without further validation; safety-critical gas or contaminant detection; evidence of mechanistic olfactory understanding; demographic representativeness of olfactory perception (the benchmark covers Western-centric science-derived descriptors); free-form or open-ended olfactory description evaluation (the benchmark uses constrained multiple-choice and multilabel formats); and individual-level psychophysical prediction (ground-truth ratings are aggregate means).",
80
- "rai:dataSocialImpact": "OP enables systematic, reproducible evaluation of LLM capability on olfaction — a sensory modality previously absent from sensory-alignment benchmarks despite its scientific and practical importance. Positive impacts include identifying capability gaps in current LLMs (e.g., the consistent SMILES-vs-name gap, the mixture-similarity failure mode, receptor knowledge idiosyncrasies) which informs research priorities, providing a baseline against which to measure progress in molecular-perceptual reasoning, surfacing cross-cultural variation in olfactory linguistic resources via the 21-language RATA evaluation, and enabling fair comparison between general-purpose LLMs and specialist olfactory models. Risks and mitigations: (i) Misuse for training: the dataset's evaluation-only design and the public release of question IDs increase the risk that questions leak into LLM training corpora over time, eroding benchmark validity. We mitigate via documentation of expected leakage trajectory and release of held-out post-cutoff RATA molecules as a memorization control. (ii) Premature claims of olfactory capability: high benchmark scores could be misinterpreted as evidence of LLM olfactory understanding suitable for fragrance, food, or safety applications; we mitigate via explicit out-of-scope documentation in this Croissant card and in the paper's limitations section. (iii) Cultural narrowness: the GS-LF/Western-perfumery descriptor lexicon could entrench a particular semantic framing of smell as the default for AI olfactory evaluation. (iv) Translation residuals: GPT-5.2-generated translations may carry subtle connotation shifts in languages without native-speaker validation; downstream users should interpret per-language results with appropriate caution. (v) Safety alignment interactions: we report cases where Claude Opus 4.6's safety filter refuses questions about hazardous odorants (e.g., nerve agents), highlighting the trade-off between safety alignment and scientific evaluation; this is documented for future benchmark designers rather than as a recommendation. The dataset is released under CC BY 4.0 to support open research and reuse with attribution.",
81
  "rai:hasSyntheticData": true,
82
  "prov:wasDerivedFrom": [
83
  {
 
73
  "multilingual",
74
  "🇺🇸 Region: US"
75
  ],
76
+ "rai:dataLimitations": "OP is an evaluation-only benchmark not intended for model training or fine-tuning. The benchmark uses standardized descriptor vocabularies and discrete response formats (multiple choice, multilabel selection from fixed lists) which enable reproducible automatic evaluation but do not capture the full richness of odor perception, including contextual effects, individual variation, free-form descriptions, or temporal/concentration dynamics. Ground-truth labels reflect agreement with established olfactory science datasets and should not be interpreted as evidence of mechanistic olfactory understanding. The RATA descriptor lexicon (138 GS-LF descriptors) and the M2OR receptor set are themselves products of specific scientific traditions and may not generalize to all olfactory perception phenomena. Multilingual RATA variants rely on automated GPT-5.2 translation; native-speaker validation was conducted for Italian, Persian, Greek, and Russian only — subtle connotation shifts may remain in the other 16 non-English languages. The benchmark is strictly not recommended for: clinical or safety-critical olfactory deployment (e.g., gas detection, food safety certification), training fragrance-design models without independent evaluation, evidence of mechanistic olfactory understanding, or commercial applications without further validation.",
77
+ "rai:dataBiases": "OP inherits the biases of its source datasets. (i) Compound coverage is drawn from industry-curated databases (GS-LF, IFRA FIG: fragrance/flavor industry odorants) and peer-reviewed academic resources (M2OR, DREAM, Leibniz-LSB@TUM); some odorants outside commercial fragrance/flavor may be underrepresented. (ii) Descriptor lexicon: the 138-descriptor RATA vocabulary follows GS-LF conventions developed primarily in English-speaking and European perfumery contexts (iii) Receptor coverage: the M2OR-derived ORA task covers human olfactory receptors that have been experimentally characterized; rare or recently discovered receptors may be underepresented. (iv) Multilingual coverage: family-level coverage is unbalanced (more Indo-European than non-Indo-European languages); GPT-5.2 translation quality reflects the model's training distribution and is expected to be stronger for higher-resourced languages.",
78
+ "rai:personalSensitiveInformation": "The dataset contains no directly Personally Identifiable Information (PII) — no names, contact details, faces, biometric data, health records, or national identifiers. The dataset does not contain personal disclosures from human subjects. Individual-level data was not collected by the authors of this work. The multilingual translation process was performed by an automated language model (GPT-5.2). No crowdsourcing was involved in benchmark construction.",
79
  "rai:dataUseCases": "OP is intended to measure factual olfactory knowledge in large language models — specifically, whether LLMs can reason from molecular information (compound names or isomeric SMILES) to perceptual properties of smell. Validated use cases include: (1) any-overlap accuracy on single-answer tasks (OC, OPD, OIn, OPl, OS, SIT) for closed-set olfactory question answering; (2) per-question multilabel F1 on RATA and ORA for descriptor and receptor knowledge evaluation; (3) Pearson correlation against human psychophysical ratings on intensity, pleasantness, and mixture similarity; (4) molecular-representation comparison via the dual-prompting (isomeric SMILES vs. compound name) design; (5) cross-lingual evaluation of olfactory descriptor knowledge via the 21-language RATA variant and ensemble approaches. Validity is established through benchmarking of 21 model configurations across 6 providers, paired statistical tests with Holm-Bonferroni correction, hierarchical bootstrap confidence intervals, and comparison to specialist models (POM ensemble, fair-retrained). Use cases NOT validated and explicitly out of scope include: training or fine-tuning olfactory prediction models (test-set integrity is required for benchmark validity); clinical or industrial olfactory deployment; fragrance or flavor product development without further validation; safety-critical gas or contaminant detection; evidence of mechanistic olfactory understanding; demographic representativeness of olfactory perception (the benchmark covers Western-centric science-derived descriptors); free-form or open-ended olfactory description evaluation (the benchmark uses constrained multiple-choice and multilabel formats); and individual-level psychophysical prediction (ground-truth ratings are aggregate means).",
80
+ "rai:dataSocialImpact": "OP enables systematic, reproducible evaluation of LLM capability on olfaction — a sensory modality previously absent from sensory-alignment benchmarks despite its scientific and practical importance. Positive impacts include identifying capability gaps in current LLMs (e.g., the consistent SMILES-vs-name gap, the mixture-similarity failure mode, receptor knowledge idiosyncrasies) which informs research priorities, providing a baseline against which to measure progress in molecular-perceptual reasoning, surfacing cross-cultural variation in olfactory linguistic resources via the 21-language RATA evaluation, and enabling fair comparison between general-purpose LLMs and specialist olfactory models. Risks and mitigations: (i) Premature claims of olfactory capability: high benchmark scores could be misinterpreted as evidence of LLM olfactory understanding suitable for fragrance, food, or safety applications; we mitigate via explicit out-of-scope documentation in this Croissant card and in the paper's limitations section. (iii) Translation residuals: GPT-5.2-generated translations may carry subtle connotation shifts in languages without native-speaker validation; downstream users should interpret per-language results with appropriate caution. The dataset is released under CC BY 4.0 to support open research and reuse with attribution.",
81
  "rai:hasSyntheticData": true,
82
  "prov:wasDerivedFrom": [
83
  {