I’m sharing English-First Prompt Lab, a small browser-local experiment developed with AI assistance to test whether translating prompts into English reduces input tokens without changing the instructions. It uses existing M2M100-418M and Qwen2.5-0.5B models, not a new base model or proof of better answers.
The tests exposed translation failures, and the small answer model also gave poor answers. Fewer tokens alone aren’t a useful result if the meaning changes. Python reference measurements and browser outputs use different runtimes, so exact translations can differ.
I’d welcome a non-sensitive example from a language you know: does the translation preserve your instructions, and do the token savings survive any corrections?
Demo and source: English-First Prompt Lab - a Hugging Face Space by vitalipavlovsk One way to make each example count as evidence rather than an anecdote: split the original prompt into its separate instructions — negations, numbers, and format requirements included, since those are where a fluent translation most often slips — have the contributor mark each one kept, changed, or dropped in the English, then fix the English until every instruction survives and recount all three versions with the answer model’s tokenizer alone: the original, the raw translation, and the corrected translation. The number worth reporting is the saving that survives correction, next to which kinds of instructions needed the fixing — if the corrected count climbs back to the original, the saving was never real for that prompt, however good the raw count looked.