Long-Form Roleplay Dataset (Work in Progress)
This is the first part of a long context roleplay dataset I've been working on for some time. I used Llama Maverick to create the first few messages, and then I used Deepseek Chat to continue extending the conversations. I originally started with pure Deepseek data, but it seemed like the conversations would degenerate into unhinged weirdness much faster. My plan is to keep on extending this dataset at least into one or two hundred turns while maintaining at least the current quality.
What This Is:
First chunk of a long-ass roleplay dataset I've been grinding on. Started by scraping a few thousand character cards from Chub, Janitor, and other places. Some of the data here is original (the raw cards), some is synthetic (me trying to make the originals less shit).
How It Was Made
First, Initial Message: Llama Maverick was used to generate the opening turns.
Second, Extending the RP: Switched to Deepseek Chat to keep the convo going. Pure Deepseek data went off the rails way faster, so I had to mix approaches.
The "User Mentality" Trick:
Biggest lesson: AI sucks at pretending to be a human RPer by default.
Fixed it by writing a script that psychoanalyzes why someone would use a given character card—what weird itch they're trying to scratch.
Example: If a bot is a "yandere gf," the script forces the AI to roleplay as a user who wants that toxic dynamic, not just some rando talking to a knife-wielding anime girl.
This made the data way fucking better, but it's still not perfect.
Current State:
Quality: Not mind-blowing, but decent. I'm trying to push these convos to 100-200 turns without them turning into nonsense.
Goal: Eventually train a small, specialized model that’s actually good at long RP.
Problems & Lessons:
- Deepseek alone = unhinged garbage after a while. Needed guardrails.
- Synthetic data is hard—AI loves to derp out without heavy guidance.
- User intent matters more than I thought. Without simulating why someone RPs a character, the dialogue feels hollow.
Want to Help?
Hit me up if you’ve got ideas for improving this dumpster fire.
- Downloads last month
- 102