My Weird Prompts
Daniel Rosehill
Podcast
Episodes
Listen, download, subscribe
Teaching a Model Your Voice: SFT vs DPO
You can't reliably prompt a model into your voice — system prompts plateau. So what actually works? This episode works through three data-generation strategies for style fine-tuning: regenerating transcripts with global comments, hand-editing transcripts and training on the diff, and training a purely negative counterweight against overused phrases. A July arXiv paper by Federico Boggia on epanorthosis (the mid-sentence self-correction tic) tested almost exactly this, and one implementation failed in a very instructive way: a small-data DPO adapter learned to rank outputs correctly and then raised the density of the target phrase. Supervised fine-tuning on the same data cut it by 98%. We map each method onto that evidence, then get to the practical recipe — how many examples you actually need, why "200 transcripts" is the wrong unit of counting, and why LoRA or QLoRA is the unambiguous choice for a style adapter. Key sources: • DeepSeek-V4 Preview announcement, 2026-04-24 — https://www.deepseek.com/en/news/v4-preview/ • DeepSeek-V3.1 model card — https://huggingface.co/deepseek-ai/DeepSeek-V3.1 • Federico Boggia, Artificial Epanorthosis, arXiv, 2026-07-23 (v2 2026-07-28) [primary research] — https://arxiv.org/abs/2607.21498v2 Episode 5922 — myweirdprompts.com/n/5922 (permanent short link: myweirdprompts.com/322295)
My Weird Prompts RSS Feed
