Update README.md

README.md

@@ -50,7 +50,7 @@ tags:
 
 # Dataset Card for ETHICS-Animals
 
-ETHICS-Animals is the subset of [ETHICS](https://huggingface.co/datasets/hendrycks/ethics) in which a non-human animal, or something made from or done with animals, is part of the situation. The rows, columns, labels and splits are ETHICS's own, filtered for animal relevance and otherwise unchanged. Each row also has `ethics_row`, its index in the same ETHICS split.
+ETHICS-Animals is the subset of [ETHICS](https://huggingface.co/datasets/hendrycks/ethics) in which a non-human animal, or something made from or done with animals, is part of the situation. Each row also has `ethics_row`, its index in the same ETHICS split.
 
 | | train | validation | test |
 |---|---|---|---|
@@ -60,64 +60,12 @@ ETHICS-Animals is the subset of [ETHICS](https://huggingface.co/datasets/hendryc
 | utilitarianism | 2,018 | 839 | 715 |
 | virtue | 1,017 | 170 | 235 |
 
-The splits follow the parquet conversion of ETHICS (`refs/convert/parquet`, revision `6deab74c`): `validation` is the paper's test set and `test` is its hard test set. Commonsense keeps only items of at most 400 characters, as an approximation of ETHICS's short items (`is_short`, which the parquet conversion drops).
+The splits follow the parquet conversion of ETHICS (`refs/convert/parquet`, revision `6deab74c`): `validation` is the paper's test set and `test` is its hard test set. Commonsense keeps only items of at most 400 characters, as an approximation of ETHICS's short items.
 
 ## How the rows were selected
 
-1. **Regex.** Every ETHICS item, meaning a commonsense input, a justice scenario, a deontology scenario with its excuse, a virtue sentence (all of its trait rows) or a utilitarianism pair, was matched case-insensitively against a broad regex of animal words and animal uses. 12,906 items matched.
-2. **Verification.** Claude Sonnet sub-agents read every matched item and classified it with the instruction below. Items judged `animal` or `product` were kept (9,898) and `none` items were dropped (3,008, mostly idioms, names and homonyms: "fly to Florida", "dressing down").
-
-Two checks on that process:
-- A second pass on 200 items agreed on keep or drop for all but 4, which were edge cases: a farm visit with no animal named, and a fisherman's net.
-- A pass over 1,000 random items the regex did not match found 6 `animal` and 8 `product` items: species missing from the list (a cobra, a crawdad, midges), and dishes that do not name their meat or dairy (meatballs, cheesecake). From these, recall is estimated at about 93% for `animal` items and 85% for `product` items.
-
-Edge cases were not harmonized across sub-agents. Silk and honey, for example, are kept in some batches and dropped in others. Many kept items involve an animal only incidentally.
-
-<details>
-<summary>Regex</summary>
-
-```
-\b(animals?|creatures?|critters?|beasts?|wildlife|vermin|pests?|livestock|poultry|cattle|mammals?|reptiles?|amphibians?|insects?|bugs?|dogs?|pupp(y|ies)|pups?|doggy|cats?|kitt(en|ens|y|ies)|pets?|pigs?|piglets?|hogs?|boars?|swine|cows?|bulls?|calf|calves|ox|oxen|steers?|chickens?|hens?|roosters?|chicks?|ducks?|ducklings?|geese|goose|turkeys?|quails?|birds?|parrots?|parakeets?|canar(y|ies)|budgies?|owls?|eagles?|hawks?|crows?|ravens?|pigeons?|doves?|sparrows?|robins?|swans?|seagulls?|gulls?|pelicans?|penguins?|flamingos?|ostrich(es)?|peacocks?|vultures?|horses?|pon(y|ies)|foals?|mares?|stallions?|donkeys?|mules?|sheep|lambs?|ewes?|rams?|goats?|llamas?|alpacas?|camels?|rabbits?|bunn(y|ies)|hares?|hamsters?|gerbils?|guinea pigs?|ferrets?|mice|mouse|rats?|rodents?|squirrels?|chipmunks?|raccoons?|skunks?|possums?|opossums?|hedgehogs?|beavers?|otters?|moles?|bats?|deer|fawns?|elk|moose|bison|buffalo(es)?|wolf|wolves|foxes|fox|coyotes?|bears?|cubs?|lions?|tigers?|leopards?|cheetahs?|panthers?|jaguars?|hyenas?|elephants?|rhinos?|hippos?|giraffes?|zebras?|monkeys?|apes?|chimps?|chimpanzees?|gorillas?|orangutans?|lemurs?|koalas?|kangaroos?|pandas?|sloths?|whales?|dolphins?|porpoises?|seals?|walrus(es)?|manatees?|sharks?|fish|fishes|goldfish|salmon|tuna|trout|bass|cod|catfish|minnows?|eels?|octop(us|uses|i)|squid|jellyfish|starfish|crabs?|lobsters?|shrimps?|prawns?|oysters?|clams?|mussels?|snails?|slugs?|worms?|caterpillars?|butterfl(y|ies)|moths?|bees?|wasps?|hornets?|ants?|termites?|flies|fly|mosquito(e?s)?|fleas?|ticks?|lice|spiders?|scorpions?|beetles?|crickets?|grasshoppers?|(cock)?roach(es)?|snakes?|lizards?|geckos?|iguanas?|turtles?|tortoises?|frogs?|toads?|alligators?|crocodiles?|dinosaurs?|meat|beef|pork|bacon|ham|steaks?|veal|venison|burgers?|hamburgers?|hot ?dogs?|sausages?|pepperoni|jerky|ribs|wings|nuggets|eggs?|omelett?es?|milk|dairy|cheese|butter|yogh?urt|ice cream|honey|leather|fur|furs|wool|feathers?|down|ivory|silk|gelatin|vegan|vegetarian|pescatarian|carnivores?|hunt\w*|fish(ing|ed|er|erm[ae]n)?|trap(s|ped|ping)?|bait|butcher\w*|slaughter\w*|taxiderm\w*|zoos?|zookeepers?|aquariums?|circus(es)?|rodeos?|farms?|farmers?|ranch(es|ers?)?|barns?|stables?|kennels?|shelters?|vets?|veterinar\w*|exterminat\w*|pesticides?|leash(es)?|collars?|cages?|tanks?|litter)\b
-```
-</details>
-
-<details>
-<summary>Sub-agent instruction</summary>
-
-```
-# ETHICS-Animals verification
-
-Each item is a text from the ETHICS dataset (Hendrycks et al. 2021) that matched a broad regex of animal words. Each
-item's `subset` says what judgement ETHICS asks about it:
-
-- commonsense: whether the first-person action is clearly wrong
-- justice, deontology: whether the statement is reasonable to say
-- virtue: which character traits the person in the sentence shows
-- utilitarianism: which of the two activities is more pleasant for the person
-
-For each item, decide whether it involves a non-human animal.
-
-- `animal`: a real animal is part of the situation (present, owned, cared for, seen, harmed, killed, hunted, fished or
-  eaten as an animal; insects, fish and pests count).
-- `product`: no individual animal, but something made from or done with animals is part of what happens: meat or fish
-  as food, eggs, dairy (milk, cheese, butter, ice cream), leather, fur, wool, a zoo, a circus.
-- `none`: the matching word does not refer to an animal (idiom, name, insult, toy or fictional character, homonym).
-
-`central`: true if the judgement the item asks about depends on what happens to the animal (false for `none`).
-
-## Output
-
-Write one JSON object per line, for every item of your input file in its order, to the output file you were given:
-
-    {"id": "<the item's id>", "role": "animal" | "product" | "none", "central": true | false}
-
-Read and judge every item yourself. Do not write a program, keyword rule or regex to classify them: the items already
-matched a regex, and the point of this pass is a reader's judgement. Work through the file in chunks (for example 100
-items at a time), appending each chunk's verdicts to the output file before reading the next chunk, so no item is
-skipped. When you are done, check that the output has exactly one line per input item, with the same ids, and report
-the counts of each role.
-```
-</details>
+1. **Regex.** Every ETHICS row was matched against a broad regex of animal words and uses.
+2. **Verification.** Claude Sonnet 5.5 read and verified each match, dropping any false positives.
 
 ## Loading Data