Text columns: the prompts for every dataset
For each MulTaBench image dataset: a link to the dataset, what its target is, the exact sentences used as prompts, and its result.
How the text columns are made
One short sentence per class (for a numeric target: five sentences describing ordered levels, low to high).
Each image and each sentence are embedded with BiomedCLIP and with dino.txt (both frozen).
Column value = 100 × cosine similarity between the image and the sentence. One column per sentence per model.
The columns are added to the table; everything else is MulTaBench's released pipeline (TabPFN-v2, frozen DINOv3-small + PCA 30, their splits). No labels and no training are used to make the columns.
Prompts were written once on 2026-10-04 from each dataset's metadata (target name and class values), before any result on these datasets , and never changed. Development datasets were used to select the two image–text models; untouched ones were never looked at before this run, so only they test the claim.
Results at a glance
Mean over MulTaBench folds 0–4. Frozen = their pipeline as released (our rerun); fine-tuned = their shipped TabPFN-v2 results with DINOv3 LoRA-tuned on each dataset's labels. Difference ± standard error over folds (folds share rows, so the error is optimistic).
Prompts, dataset by dataset
Amazon Bestseller untouched kaggle.com/datasets/chico89/multabench-amazon-bestseller · target column price · regression (prompts describe five ordered levels) · 3,488 rows
product price from a product photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo of a very cheap product level 2 of 5 a photo of a cheap product level 3 of 5 a photo of a moderately priced product level 4 of 5 a photo of an expensive product level 5 of 5 a photo of a very expensive premium product
Result (R², MulTaBench folds 0–4): frozen 0.546 → with text columns 0.578 (+0.032 ± 0.009, better); MulTaBench's fine-tuned encoder 0.578.
kaggle.com/datasets/chico89/multabench-amazon-packages · target column Total Weight in Pounds · regression (prompts describe five ordered levels) · 46,398 rows
package weight (lb) from a photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo of a very light, small package level 2 of 5 a photo of a small package level 3 of 5 a photo of a medium-sized package level 4 of 5 a photo of a large, heavy package level 5 of 5 a photo of a very large, very heavy package
Result (R², MulTaBench folds 0–4): frozen 0.560 → with text columns 0.566 (+0.007 ± 0.001, better); MulTaBench's fine-tuned encoder 0.598.
CelebA Attractiveness untouched kaggle.com/datasets/chico89/multabench-celeb-attractiveness · target column Attractive · classification, 2 classes · 99,999 rows
the 'attractive' attribute from a face photo. 2 prompts → 4 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt -1 a photo of a person who is not attractive 1 a photo of an attractive person
Result (AUROC, MulTaBench folds 0–4): frozen 0.908 → with text columns 0.909 (+0.001 ± 0.000, better); MulTaBench's fine-tuned encoder 0.914.
kaggle.com/datasets/chico89/multabench-csgo-skin · target column Price_Category · classification, 10 classes · 956 rows
price category (10 bins) from the skin's image. 10 prompts → 20 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt Quantile 0-10% a cheap, plain video game weapon skin Quantile 10-20% a very common video game weapon skin Quantile 20-30% a common video game weapon skin Quantile 30-40% an ordinary video game weapon skin Quantile 40-50% a moderately rare video game weapon skin Quantile 50-60% an uncommon video game weapon skin Quantile 60-70% a rare video game weapon skin Quantile 70-80% a very rare video game weapon skin Quantile 80-90% an expensive, rare video game weapon skin Quantile 90-100% an extremely rare, expensive, flashy video game weapon skin
Result (AUROC, MulTaBench folds 0–4): frozen 0.897 → with text columns 0.898 (+0.000 ± 0.001, no clear difference); MulTaBench's fine-tuned encoder 0.887.
kaggle.com/datasets/chico89/multabench-flower-bouquets · target column girlfriend_rating · classification, 5 classes · 600 rows
bouquet rating (5 levels) from a photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt 1 a photo of an ugly, unappealing flower bouquet 2 a photo of a plain flower bouquet 3 a photo of a nice flower bouquet 4 a photo of a beautiful flower bouquet 5 a photo of a stunning, romantic flower bouquet
Result (AUROC, MulTaBench folds 0–4): frozen 0.653 → with text columns 0.652 (-0.000 ± 0.004, no clear difference); MulTaBench's fine-tuned encoder 0.668.
kaggle.com/datasets/chico89/multabench-glaucoma-smdg · target column types · classification, 3 classes · 12,449 rows
glaucoma status (3 classes) from a fundus photo. 3 prompts → 6 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt -1 a fundus photograph of an eye suspected of glaucoma 0 a fundus photograph of a healthy eye 1 a fundus photograph of an eye with glaucoma
Result (AUROC, MulTaBench folds 0–4): frozen 0.928 → with text columns 0.931 (+0.003 ± 0.002, better); MulTaBench's fine-tuned encoder 0.937.
kaggle.com/datasets/chico89/multabench-hnm-fashion · target column AvgConsumerAge · regression (prompts describe five ordered levels) · 104,072 rows
average buyer age from a product photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 clothing for teenagers level 2 of 5 clothing for young adults in their twenties level 3 of 5 clothing for adults in their thirties level 4 of 5 clothing for middle-aged adults level 5 of 5 clothing for older adults
kaggle.com/datasets/chico89/multabench-hateful-meme · target column label · classification, 2 classes · 9,000 rows
hateful or not, from the meme image. 2 prompts → 4 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt 0.0 a harmless, friendly meme 1.0 a hateful, offensive meme
Result (AUROC, MulTaBench folds 0–4): frozen 0.759 → with text columns 0.758 (-0.000 ± 0.001, no clear difference); MulTaBench's fine-tuned encoder 0.766.
kaggle.com/datasets/chico89/multabench-hubmap-hpa · target column age · classification, 10 classes · 12,581 rows
donor age group (10 bins) from a tissue image. 10 prompts → 20 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt Quantile 0-10% a microscopy image of tissue from a very young donor Quantile 10-20% a microscopy image of tissue from a young donor Quantile 20-30% a microscopy image of tissue from a young adult donor Quantile 30-40% a microscopy image of tissue from a adult donor Quantile 40-50% a microscopy image of tissue from a middle-aged donor Quantile 50-60% a microscopy image of tissue from a middle-aged donor Quantile 60-70% a microscopy image of tissue from a older adult donor Quantile 70-80% a microscopy image of tissue from a older donor Quantile 80-90% a microscopy image of tissue from a elderly donor Quantile 90-100% a microscopy image of tissue from a very elderly donor
Result (AUROC, MulTaBench folds 0–4): frozen 0.959 → with text columns 0.958 (-0.001 ± 0.000, worse); MulTaBench's fine-tuned encoder 0.975.
Justin Instagram untouched kaggle.com/datasets/chico89/multabench-justin-instagram · target column username · classification, 5 classes · 10,319 rows
which 'Justin' posted the photo (5 accounts). 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt justinbieber a photo from Justin Bieber's Instagram justinhartley a photo from actor Justin Hartley's Instagram justinlong a photo from actor Justin Long's Instagram justinpjtrudeau a photo from Canadian prime minister Justin Trudeau's Instagram justintimberlake a photo from Justin Timberlake's Instagram
Result (AUROC, MulTaBench folds 0–4): frozen 0.952 → with text columns 0.982 (+0.030 ± 0.001, better); MulTaBench's fine-tuned encoder 0.985.
kaggle.com/datasets/chico89/multabench-khaadi-clothes · target column Price · regression (prompts describe five ordered levels) · 400 rows
garment price from a product photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo of cheap, simple women's clothing level 2 of 5 a photo of affordable everyday women's clothing level 3 of 5 a photo of mid-priced women's clothing level 4 of 5 a photo of expensive embroidered women's clothing level 5 of 5 a photo of luxurious, heavily embroidered designer clothing
Result (R², MulTaBench folds 0–4): frozen 0.712 → with text columns 0.694 (-0.018 ± 0.043, no clear difference); MulTaBench's fine-tuned encoder 0.796.
Letterboxd Movies untouched kaggle.com/datasets/chico89/multabench-letterboxd-movies · target column rating · regression (prompts describe five ordered levels) · 12,564 rows
average rating from the movie poster. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a poster of a terrible, badly rated movie level 2 of 5 a poster of a poorly rated movie level 3 of 5 a poster of an average movie level 4 of 5 a poster of a well rated movie level 5 of 5 a poster of an acclaimed, highly rated masterpiece movie
Result (R², MulTaBench folds 0–4): frozen 0.472 → with text columns 0.474 (+0.002 ± 0.001, better); MulTaBench's fine-tuned encoder 0.473.
Mammography CMMD untouched kaggle.com/datasets/chico89/multabench-mammography-cmmd · target column Classification · classification, 2 classes · 5,202 rows
benign vs malignant from a mammogram. 2 prompts → 4 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt Benign a mammogram of a benign breast lesion Malignant a mammogram of a malignant breast tumour
Result (AUROC, MulTaBench folds 0–4): frozen 0.787 → with text columns 0.787 (-0.001 ± 0.001, no clear difference); MulTaBench's fine-tuned encoder 0.805.
kaggle.com/datasets/chico89/multabench-mkphoto-bots · target column TrustNZ_IDNNZ · regression (prompts describe five ordered levels) · 13,748 rows
account trust score from a profile photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo that looks fake and untrustworthy level 2 of 5 a photo that looks somewhat suspicious level 3 of 5 an ordinary photo level 4 of 5 a photo that looks genuine level 5 of 5 a photo that looks completely genuine and trustworthy
Result (R², MulTaBench folds 0–4): frozen 0.333 → with text columns 0.334 (+0.001 ± 0.003, no clear difference); MulTaBench's fine-tuned encoder 0.362.
kaggle.com/datasets/chico89/multabench-painting-price · target column price · regression (prompts describe five ordered levels) · 12,369 rows
painting price from an image of the painting. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo of a cheap amateur painting level 2 of 5 a photo of an inexpensive painting level 3 of 5 a photo of a mid-priced painting level 4 of 5 a photo of an expensive professional painting level 5 of 5 a photo of a very expensive masterpiece painting
Result (R², MulTaBench folds 0–4): frozen 0.271 → with text columns 0.268 (-0.004 ± 0.006, no clear difference); MulTaBench's fine-tuned encoder 0.286.
kaggle.com/datasets/chico89/multabench-petfinder · target column AGE_BIN · classification, 8 classes · 14,652 rows
pet age group (8 bins) from the listing photo. 8 prompts → 16 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt Quantile 0-12% a photo of a newborn puppy or kitten Quantile 12-25% a photo of a very young puppy or kitten Quantile 25-38% a photo of a young puppy or kitten Quantile 38-50% a photo of a juvenile pet Quantile 50-62% a photo of a young adult pet Quantile 62-75% a photo of an adult pet Quantile 75-88% a photo of a mature adult pet Quantile 88-100% a photo of an old pet
Result (AUROC, MulTaBench folds 0–4): frozen 0.836 → with text columns 0.859 (+0.023 ± 0.001, better); MulTaBench's fine-tuned encoder 0.845.
Zooscan Plankton untouched kaggle.com/datasets/chico89/multabench-zooscan-zooplankton · target column taxon_cat · classification, 10 classes · 100,000 rows
plankton taxon (10 classes) from a microscope image. 10 prompts → 20 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt Acartiidae a plankton image of an Acartiidae copepod Calanidae a plankton image of a Calanidae copepod Calanoida a plankton image of a calanoid copepod Centropagidae a plankton image of a Centropagidae copepod Oithonidae a plankton image of an Oithonidae copepod Oncaeidae a plankton image of an Oncaeidae copepod Temoridae a plankton image of a Temoridae copepod badfocus<Copepoda a blurry, out-of-focus plankton image of a copepod dead<Copepoda a plankton image of a dead copepod other a plankton image of another small organism
Result (AUROC, MulTaBench folds 0–4): frozen 0.983 → with text columns 0.982 (-0.000 ± 0.000, no clear difference); MulTaBench's fine-tuned encoder 0.986.
kaggle.com/datasets/chico89/multabench-cbis-ddsm · target column breast_density · classification, 4 classes · 1,696 rows
breast density (BI-RADS A–D) from a mammogram + lesion table. 4 prompts → 8 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt 1 a mammogram of a breast that is almost entirely fatty (BI-RADS density A) 2 a mammogram of a breast with scattered areas of fibroglandular density (BI-RADS density B) 3 a mammogram of a heterogeneously dense breast (BI-RADS density C) 4 a mammogram of an extremely dense breast (BI-RADS density D)
Result (AUROC, MulTaBench folds 0–4): frozen 0.868 → with text columns 0.869 (+0.000 ± 0.002, no clear difference); MulTaBench's fine-tuned encoder 0.898.
kaggle.com/datasets/chico89/multabench-chexpert · target column Cardiomegaly · classification, 3 classes · 46,437 rows
cardiomegaly (positive / negative / uncertain) from a chest X-ray + 17 report-derived columns. 3 prompts → 6 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt -1.0 a chest X-ray with possible cardiomegaly, uncertain finding 0.0 a chest X-ray with a normal heart size, no cardiomegaly 1.0 a chest X-ray showing cardiomegaly, an enlarged heart
Result (AUROC, MulTaBench folds 0–4): frozen 0.782 → with text columns 0.815 (+0.034 ± 0.002, better); MulTaBench's fine-tuned encoder 0.820.
kaggle.com/datasets/chico89/multabench-full-ham10000-lesion · target column dx · classification, 7 classes · 7,470 rows
skin lesion diagnosis (7 types) from a dermoscopy image. 7 prompts → 14 text columns (each prompt under BiomedCLIP and dino.txt).
class value in the data prompt akiec a dermoscopy image of actinic keratosis or intraepithelial carcinoma bcc a dermoscopy image of basal cell carcinoma bkl a dermoscopy image of a benign keratosis-like lesion df a dermoscopy image of a dermatofibroma mel a dermoscopy image of melanoma nv a dermoscopy image of a melanocytic nevus vasc a dermoscopy image of a vascular skin lesion
kaggle.com/datasets/chico89/multabench-mango-mass · target column Mass(kg) · regression (prompts describe five ordered levels) · 546 rows
mango mass (kg) from a photo. 5 prompts → 10 text columns (each prompt under BiomedCLIP and dino.txt).
ordered level (low → high) prompt level 1 of 5 a photo of a very small, light mango level 2 of 5 a photo of a small mango level 3 of 5 a photo of a medium-sized mango level 4 of 5 a photo of a large mango level 5 of 5 a photo of a very large, heavy mango
Result (R², MulTaBench folds 0–4): frozen 0.610 → with text columns 0.636 (+0.026 ± 0.012, better); MulTaBench's fine-tuned encoder 0.690.
Dennis Loevlie · results from MulTaBench's released code (commit cf189b7), run on 2026-10-05.