This hobbyist fine-tuned Qwen on a 16 GB GPU to answer without ever using the letter 'e'
Fine-tuned Qwen on my 16gb GPU to respond without the letter e
Kevin Wu trained Qwen3-1.7B to write lipograms — everyday answers that avoid the letter 'e' — on a single 16 GB GPU. Prompting alone failed completely, and a token mask that blocks 'e' got usable answers to 26% but produced nonsense. Supervised fine-tuning on 180 hand-built 'e'-free examples reached 35%, then RL with an LLM judge as the reward pushed it further, teaching the model to answer coherently while dodging the letter.
They read text in tokens, and not words, so whether something contains an 'e' is something the model cannot directly see.
- jamesdutc
One year in school, we were made to write essays in E-Prime: https://en.wikipedia.org/wiki/E-Prime
If I remember correctly, we would have one pointed deducted for every instance we failed to eliminate.
I suspect this would be nearly trivial to get train (or get an existing model) to do, but, like writing in cursive, one could argue that these activities (as absurd as they may seem…) genuinely and positively contribute to a child's growth and development.
(In fact, I'm actually slowly reteaching myself a proper cursive…)
- _ache_
Qwn is, I think, an à propos alias.
- NooneAtAll5
I wonder if instead of outright masking bad tokens, you could've done the training while slowly increasing negative weights on those