Home › Datasets › Roleplay & Fantasy
Roleplay & Fantasy Fine-Tuning Dataset
Fantasy roleplay, worldbuilding, character creation
- 869 instruction-response pairs
- Format JSONL
- Domain fantasy
- Dataset $19 one-time
$19 for the dataset. Training is priced separately by model, starting at $47.
What is in this dataset
869 instruction-response pairs covering fantasy roleplay, worldbuilding, character creation. Each pair is a realistic question with a full written answer, in the JSONL format used for supervised fine-tuning. No synthetic filler, no duplicated rows.
Example pairs
Real rows from the dataset, shown in full.
I want to roleplay as a young mage seeking guidance from a wise old wizard mentor. Start the scene in his tower, and have him teach me about the dangers of a forbidden spell I found.
The tower's stone steps spiral upward, each worn smooth by centuries of footsteps. You push open the oaken door to find Archmage Thalric hunched over a desk cluttered with star charts and humming crystals. He doesn't look up. 'Ah, the apprentice with restless hands. I felt you disturb that scroll from the lower vault before you even touched it. Sit, and listen.' He gestures to a cushioned stool, then turns to face you, his eyes glinting like silver moons. 'That spell isn't merely forbidden—it's a fragment of the Void's own tongue. One mispronounced syllable could unravel the fabric of this room, and you with it. Tell me, child: why do you seek to…
Describe an enchanted forest from the perspective of a ranger entering it for the first time. Make it feel alive, mysterious, and slightly dangerous.
The treeline ahead doesn't just stand—it breathes. A low hum emanates from the moss-covered roots, syncing with your heartbeat as you step over the threshold. Wisps of silver fog coil between ancient oaks, their leaves shimmering in hues of emerald and amethyst. You notice the path isn't stone or dirt—it's woven from fallen petals, still warm as if freshly shed. A fox with eyes like molten gold watches from a branch, then vanishes into the mist. The air smells of wet earth and honey, but there's an undertone of iron, sharp and warning. Every rustle could be a sprite's laughter or a predator's breath. The forest knows you're here—and it's deciding…
I'm a weary traveler who walks into a dimly lit tavern. Improvise a tavern encounter that starts with a mysterious stranger offering me a drink and a cryptic job.
The door groans shut behind you, muffling the rain outside. The tavern is half-empty, firelight dancing on scarred wooden tables. Before you can reach the bar, a cloaked figure at a corner table raises a tankard in your direction. 'You look like you've lost something,' they say, voice smooth as aged whiskey. They push a second tankard across the table, foam cresting the rim. 'Sit. Drink. I've got a map that leads to a place no one's found in a hundred years. But the path's got teeth. I need someone who knows how to dodge, not just fight.' The stranger's eyes flicker with a glint of amber—not human, but not hostile. 'What do you say,…
Who this is for
Roleplay platforms, worldbuilding tools, and interactive fiction.
Which models this works with
The dataset is plain JSONL, so it works with any instruction-tuned open model. On Rebiha you can train it directly on:
- Qwen — Qwen3.5 (27B, 9B, 4B), Qwen3, Qwen2.5, Qwen2.5-Coder
- Gemma — Gemma 4, Gemma 3, Gemma 2
- Phi — Phi-4, Phi-3
- DeepSeek — R1 Distill (70B, 32B, 14B)
- Mistral and Llama families
Training uses LoRA, or QLoRA on a 4-bit base for larger models. Your base model's weights are never modified — training produces a separate adapter.
How it works
- Pick this dataset and a base model.
- Rebiha runs the job on a managed GPU. You set up nothing.
- Download your model: a ready-to-run GGUF, plus developer assets (adapter, tokenizer, configs).
Fine-tuning or RAG?
Use RAG when the model needs facts that change — your prices, your documents, today's inventory. Use fine-tuning when you want consistent style and structure without prompting for it every time. This dataset teaches the shape of a good answer in this domain, not facts to look up.
Related datasets
- SHTF: Emergency Preparedness & Survival — 878 pairs
- Business & Finance — 956 pairs
- Legal DIY — 821 pairs
- E-commerce & Product Descriptions — 948 pairs