I checked Hugging Face for a neuroscience\/biology QA dataset that wasn’t either scraped Wikipedia with zero voice or dumbed down to the point of being wrong. Nothing fit, so I built one.<\/p>\n
NeuroBio GenZ 1K<\/strong>: about 1000 questions like “why does doomscrolling give me dopamine” or “why do I forget why I walked into a room,” each answered in 2-5 sentences, casual tone, but using the actual terminology instead of vague placeholders. Nucleus accumbens instead of “reward center.” GABA instead of “chill chemical.” The system prompt explicitly banned the “dopamine is the pleasure chemical” style oversimplification that’s everywhere in this space.<\/p>\n Fully local generation. Ornith 1.5 35B-A3B, Q3_K_XL quant, through llama.cpp on a Ryzen 7 5700X, 32GB DDR4, 8GB GPU. No API calls, no cloud spend.<\/p>\n The part I actually want to share because it’s a real lesson: I ran the whole generation at 10-13 tok\/s because I hadn’t figured out the right server flags yet. After the run finished I found the config that gets the exact same hardware to 30-35 tok\/s: speculative decoding via Dataset: https:\/\/huggingface.co\/datasets\/luka0x12\/NeuroBio-GenZ-1K<\/a> CC BY 4.0, use it for whatever.<\/p>\n<\/div>\n submitted by \/u\/luka_0x12 <\/a> --spec-type draft-mtp<\/code> plus properly tuned --n-cpu-moe<\/code> for MoE expert offload. Both configs are in the repo if you’re doing something similar and don’t want to waste the time I did.<\/p>\n
Source code + server scripts: https:\/\/github.com\/luka0x12\/NeuroBio-GenZ-1K<\/a><\/p>\n
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
