Nicolay (nicolay)

Treat finetuning as an experiment! It's cheap to train multiple models, so evaluate them on your own test sets & pick the best. Start with fast evals (LLM as a judge,…) before investing in production-grade testing.

• 0 replies
• 0 recasts
• 0 reactions

When building finetuning datasets, don't cherry-pick only the "good" examples! Relabel & fix bad ones to cover your full input space. A bit of noise is OK as long as examples are correct on average.

• 0 replies
• 0 recasts
• 0 reactions

Use GPT-4 to go from idea to working prototype, then finetune a smaller model to scale cost-effectively. Finetuning can dramatically reduce costs & latency while maintaining quality.

• 0 replies
• 0 recasts
• 0 reactions

The more specific your task and the more it diverges from being a general purpose chatbot, the more likely you are to get good results from finetuning vs prompting alone. With the caveat, if large models completely fail, even a finetuned model will likely fail.

• 0 replies
• 1 recast
• 0 reactions

For low-resource languages, finetuning a model to generate outputs in that language may be challenging due to lack of data. But finetuning to understand the language while outputting English is more feasible.

• 0 replies
• 0 recasts
• 0 reactions

When building LLM apps, default to prompts! Only finetune if absolutely needed for quality, speed or cost. Iterate fast with prompts + few-shot/RAG before investing in finetuning.

• 0 replies
• 0 recasts
• 1 reaction

Finetuning 101: First, write an excellent prompt to establish a baseline & prove your task is possible. A great prompt is a strong predictor that finetuning will improve results further.

• 0 replies
• 0 recasts
• 0 reactions