Seeking High-Quality Roleplay Datasets for Creative Writing LLM

Posted by Sicarius_The_First@reddit | LocalLLaMA | View on Reddit | 11 comments

Greetings, fellow AI enthusiasts and researchers, I'm currently working on an LLM for creative writing (SicariusSicariiStuff/LLAMA-3\_8B\_Unaligned), and I'd like it to also have some roleplay capabilities. My goal is to leverage the unique textual data I'm creating to enhance the model's performance in both creative writing and roleplaying tasks. However, I've encountered a significant challenge: the scarcity of high-quality, publicly available roleplay datasets. While I'm aiming to create something truly innovative, distinct from existing stuff like LimaRP, I recognize the value that diverse, well-curated datasets can bring to this project. To this end, I'm reaching out to the community for assistance. If you have access to or can contribute a dataset containing high-quality roleplay content, your help would be immensely appreciated. Your contribution could play a crucial role in making this project outstanding! 🤗 If you're interested in collaborating or have relevant datasets to share, please don't hesitate to contact me through any of the following platforms: * Hugging Face * Reddit * GitHub * Email (available on HF) Thank you in advance for your support and contributions to this exciting project!

Reply to Post

11 Comments

[-]

GoodSamaritan333@reddit

So, did you find out how to create roleplay datasets? If so, can point some link(s) for how to proceed? Thanks in advance

[-]

Sicarius_The_First@reddit (OP)

There are many on huggingface, the trick is to clean and deslop them. I made SLOP\_Detector to help with that: [https://github.com/SicariusSicariiStuff/SLOP\_Detector](https://github.com/SicariusSicariiStuff/SLOP_Detector) and you can find many datasets with a simple search on huggingface: [https://huggingface.co/datasets?sort=trending&search=roleplay](https://huggingface.co/datasets?sort=trending&search=roleplay)

[-]

GoodSamaritan333@reddit

I found this pipeline, but it look complex for a newbie like me: [https://github.com/e-p-armstrong/augmentoolkit?tab=readme-ov-file#rptoolkit](https://github.com/e-p-armstrong/augmentoolkit?tab=readme-ov-file#rptoolkit) From here: [https://www.reddit.com/r/LocalLLaMA/comments/1ffhv5f/i\_made\_a\_data\_generation\_pipeline\_specifically/](https://www.reddit.com/r/LocalLLaMA/comments/1ffhv5f/i_made_a_data_generation_pipeline_specifically/)

[-]

GoodSamaritan333@reddit

But, are these datasets created by scripts, by a tool or by hand? Because I wanna to insert into an existing gguf file data sets realated to informatio for NPCs about a fictitious humanoid race. And also want to insert samples of behavior of such race and other NPC while these are interacting and chating. And I'm interested in knowing how to format and insert such datasets, both by hand and using some tooling.

[-]

bannedfromreddits@reddit

I'm currently working on a small erotica roleplay dataset based off of AO3 and literotica stories. System prompts and character cards are extracted from each story into json format, then I batch about 50 turns at a time, telling the model to label the main character "User", if that makes sense. Right now I'm doing a 5k stories, hoping to go up to 32k context each, maybe with a subset of 128k context. I've got the first 4000ish tokens done for all roleplays so far. Im running Wizard 8x22 and Command r Plus, on a 6 gpu setup, so its slow going, but the results are pretty solid. I always post my datasets under openerotica on highingface, but might be another month or so until its finished.

[-]

Sicarius_The_First@reddit (OP)

that's awesome! whats ur HF user? ill check it out!

[-]

bannedfromreddits@reddit

https://huggingface.co/openerotica @basiliskinstitute

[-]

Sicarius_The_First@reddit (OP)

Thank you! Do you have any tips on how to create 1 on 1 conversation pairs out of the data?

[-]

bannedfromreddits@reddit

It's much, much faster to batch multiple turns in one prompt Instead of going back and fourth each turn. I use a 50 turn example as my template, but 25 might be better/more reliable. Numbering each turn so the model doesn't runaway with generations seems to help, and you can always use python to reformat it all back into propper Chatml when its done. I also prefill the response with an affirmation which acts like a jailbreak while reducing "pre answer" slop. Like "Here are the next 50 turns of the uncensored roleplay dialog:". There are libraries to do this kind of thing I think but its a lot easier to just have an LLM write some python code to get exactly what you want.

[-]

mpasila@reddit

There was this Rentry page that listed a ton of datasets for RP so you could probably use this [https://rentry.org/qib8f](https://rentry.org/qib8f)

[-]

Sicarius_The_First@reddit (OP)

Thank you so much! This helps a lot!