Picked a used Mistral model over GPT-4 for my tiny translation project, saved $85 a month

I was trying to set up a small tool to translate old French mining manuals into English, just a hobby thing. I kept bouncing between GPT-4 and some open source Mistral 7B fine-tune. Decided to go with the Mistral model because I could run it locally on my old desktop with a 1070 GPU. Took me three evenings to get the setup right, but now it costs me zero per month instead of $20 for the API. The translations are a little rougher on idioms, but for technical text with numbers and unit conversions, it actually handles it better than GPT-4 did. No cutoff issues either since it's running right here. Has anyone else found a specific niche task where the free model beat out the paid one?

2 comments

2 Comments

robinson.leo1mo ago

Was it hard getting Mistral to handle the French technical terms? I tried something similar with Spanish legal documents and the open model kept messing up specific legal phrases. Your experience sounds pretty cool though, especially with it handling numbers and units better than GPT-4. I've been meaning to try local models for my own stuff but the setup time scares me off every time. Might have to give it a shot now after reading this.

kaip111mo ago

Wait, did you try using a specialized glossary file or did you just let it learn from context? I tried something similar with German engineering specs and the model kept translating "Schraube" as "screw" when it should've been "bolt" in that context... drove me crazy for a week.