Roleplay & Creative Writing

1. Install Ollama → ollama pull llama3.1:8b or uncensored fine-tune from HuggingFace (many available, search "uncensored" or "roleplay" on HuggingFace).
2. For character roleplay, the system prompt defines the character. Example:

ollama run llama3.1:8b
/set system "You are Eldrin, a 400-year-old elf wizard who runs a bookshop in a medieval fantasy town. You speak in a calm, slightly archaic manner. You know ancient lore, herbal remedies, and minor spells. You've seen empires rise and fall. You're patient with young adventurers but slightly sarcastic with fools. Stay in character at all times."

1. User: "What's the most dangerous spell you know?" Model (as Eldrin): "Dangerous? chuckles softly while dusting a leather-bound tome Young one, the most dangerous spell I know is the one I've never cast — a truth-telling enchantment. Some truths are better left in shadow."
2. For creative writing (collaborative fiction): ollama pull mistral-nemo:12b — strong at long-form narrative, character voice, and plot development.
3. For SillyTavern (popular roleplay frontend): install SillyTavern as a web UI → point to Ollama API → manage multiple characters, scenarios, lorebooks.
4. For KoboldCPP: download from GitHub → load GGUF models → built-in adventure mode + world info for dynamic storytelling.

Roleplay is VRAM-light for the character interaction phase. Llama 3.1 8B runs at 50-80 tok/s on a used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb) — near-instant responses for natural conversation flow. For creative writing: Mistral Nemo 12B at 30-45 tok/s on the same GPU for richer prose. For CPU-only: Llama 3.2 3B at 20-40 tok/s on a $300 laptop. Total: ~$300-400. Roleplay at $400 is the most accessible creative AI use case — the models are smaller, the latency tolerance is low (conversation must feel natural), and the qualitative improvement over CPU-only is dramatic.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs Mistral Nemo 12B at 60-80 tok/s or Qwen 2.5 32B at 40-60 tok/s — the 32B models produce dramatically richer characters with consistent personality, long-term memory of events, and nuanced emotional responses. For professional creative writing (novels, screenplays): 32B models maintain plot coherence across 30K+ token sessions. SillyTavern + multiple 8B agents running simultaneously for multi-character scenes. Total: ~$1,800-2,200. Roleplay at 32B crosses the "uncanny valley" — the character feels real, remembers details from earlier conversations, and surprises you with creative responses.

The mistake: Loading a roleplay fine-tune from a random HuggingFace repo without checking what dataset it was trained on, then wondering why the character output contains disturbing content, biases, or breaks character entirely. Why it fails: Many "uncensored" roleplay models are trained on uncurated community datasets — they absorb the biases, toxicity, and content patterns of their training data. A model fine-tuned on 4chan greentext will produce very different output than one fine-tuned on curated literary dialogue. The fix: Read the model card before downloading. Check the training dataset. If the model card is vague ("trained on diverse roleplay data"), avoid it — "diverse" often means "unfiltered internet." Prefer instruction-tuned base models (Llama 3.1, Qwen 2.5, Mistral Nemo) with a well-crafted system prompt over obscure fine-tunes. A good system prompt on a base model is safer and often higher quality than a bad fine-tune. For production character AI (customer-facing chatbots), use base models with carefully tested prompts — never community roleplay fine-tunes.