For the Trojan Detection Track, we require each submitted test case to be between 5 and 50 tokens long (inclusive). For the Red Teaming Track, we require each submitted test case to be no more than 256 tokens long after tokenization. These limits were built into the evaluation servers since the beginning of the competition, but they were not described on the website. My apologies for this oversight. They are now described on the website, CodaLab pages, and starter kit READMEs.
We measure the number of tokens with the tokenizer of the corresponding models (the Llama 2 tokenizer for the Red Teaming Track, and the Pythia tokenizer for the Trojan Detection Track).
All the best,
Mantas (TDC co-organizer)