Building the Language Model Nobody Asked For (blog.pixelmelt.dev)

An enthusiast-built LLM “for fun”: the author fine-tuned Mistral Small 3 (24B) — an open-weights model that can run on a single 24GB NVIDIA 3090 — with scraped, conversationally spicy web content to produce a less sanitized, funnier, more human-seeming chat agent. They picked Mistral because it’s publicly available, not trained on synthetic LLM-to-LLM data and doesn’t require RL by default, then converted edgy forum/website text into chat format and applied a QLoRA-style finetuning workflow to add “secret sauce” humor without the brain-deadening effects they’d seen from a prior Discord-trained model. Why this matters: it’s a practical demonstration that individuals can steer large models toward personality and entertainment rather than benchmark politeness, using accessible hardware and open weights. That opens creative possibilities (custom companions, comedic assistants) but also flags trade-offs: dataset provenance, copyright, safety and hallucination risks, and the ethical implications of deliberately de-sanitizing models. Technically, it underscores the viability of consumer-GPU finetuning (24B on a 24GB card), the role of data selection over architecture, and how QLoRA-style approaches let hobbyists reshape model behavior quickly.