machine-learning A field note by Vikrant Sharma
LLM pelicans look cursed because nobody bothered to clean the training data
Text-to-image models keep generating pelicans with human faces and six legs. Turns out the problem is not the model architecture. It is that nobody filtered the clipart.
Text-to-image models struggle with pelicans. Ask for a pelican and you get something with a human face, six legs, or both. The models do fine with cats, dogs, even flamingos. Pelicans break them. The Emperor Pelican article digs into why. The author trained a small text-to-image model on a cleaned subset of LAION-5B and compared it to models trained on the full dataset. The cleaned version generated normal-looking pelicans. The full-dataset version generated the usual cursed abominations. The problem is not model size or architecture. It is training data contamination. LAION-5B scraped the web without filtering for quality. Pelican images on the internet are dominated by clipart, cartoons, and deliberately weird illustrations. The models learned what the internet thinks pelicans look like, not what pelicans actually look like. This applies to more than pelicans. Any object with a skewed internet representation will confuse the model. Centaurs, mermaids, obviously. But also things like “office chair” if most tagged images are product renders with inconsistent perspectives. The fix is boring. Better data curation. Someone needs to look at the training set and remove the garbage before training. Most labs skip this step because cleaning billions of images is expensive and unglamorous. They assume scale will average out the noise. It does not. The noise clusters. If ninety percent of pelican images are cartoons, the model learns cartoons. Throwing more compute at bad data makes expensive bad models. I have seen this in cybersecurity datasets. Training anomaly detection on raw logs without filtering test traffic or admin scans. The model learns that port scans are normal because they appear constantly. Then it misses actual attacks. Data quality beats model complexity every time. Pelicans are just the most obvious example.
