Finally got my tiny language model to stop giving nonsense answers

I had to pick between adding way more training data or just tweaking the prompt structure a lot. Went with the prompt tweaks, and after about 20 tries, it started giving coherent replies on my test set. Anyone else get stuck on something simple like that?

2 comments

Log in to join the discussion

Log In

2 Comments

stone.jesse5mo ago

Honestly, skipping the data grind sounds risky. A better prompt is just a band-aid if the model's foundation is shaky.

8

miam115mo ago

Man I remember reading somewhere that prompt structure can matter more than people give it credit for. It's wild how just rephrasing a few lines can make a model go from gibberish to actually useful. Sounds like you saved yourself a ton of time skipping the data grind.

0