What is the best prompting strategy for DeepSeek V4 Flash?

Question

Honestly, I am super fed up with DeepSeek V4 Flash right now. Ive been trying to get it to summarize these long legal docs for my firm, but it just keeps hallucinating details or skipping entire sections. My logic was that a flash model would be fast enough for my 5pm deadlines, but it is just wasting my time. I have tried chain of thought, giving it examples, even tried system prompts but nothing seems to stick. Is there some specific prompting strategy I am missing here or is this model just totally useless for structured work? I have a budget to burn but im about ready to switch back to Claude if this keeps up...

yxmyfkpizm · Accepted Answer

Honestly, flash models are pretty hit or miss for heavy legal stuff. You should try breaking the documents into smaller chunks rather than feeding the whole file at once. Also, force the output into a specific schema like JSON to prevent it from wandering off. If you need rock solid accuracy, maybe just run a quick test with OpenAI GPT-4o 128k context window instead.

JamesDox · Answer

Regarding what #2 said about sticking to Claude, they're not wrong, but if you're stuck with DeepSeek, try setting the temperature to 0. It stops the model from getting too creative with legal jargon. Also, use OpenAI GPT-4o 128k context if you need better reasoning. Flash models are just built for speed, not accuracy. Maybe use a local tool for the initial extraction, then polish it up. It saves a ton of headaches.

GlastonburyGoer · Answer

Had a moment to think on this. Stick to Anthropic Claude 3.5 Sonnet for legal accuracy. Flash models hallucinate too much for contracts. Be careful with those 5pm deadlines tbh.