|
|
--- |
|
|
language: |
|
|
- en |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
|
|
|
extra_gated_prompt: "This Dataset contains sexual explicit material. Please confirm your age by checking the box below." |
|
|
extra_gated_button_content: "Age Restriction" |
|
|
extra_gated_fields: |
|
|
I confirm that I am over the age of 18: checkbox |
|
|
--- |
|
|
Dataset ERP dataset generated from deepseek R1. This was a bit intensive. I write a python scrip that would generate a random plot, the ERP theme of which was chosing at random from about 20 different types of themes (One of the more tame ones, for example, "Romance Erotica"). I would then guide the inference by using the "user" role to switch perspectives between the two characters, and continue the story. |
|
|
|
|
|
I then wrote another script that used deepseek-cat to fix the reasoning portions. Since I my instructions were automated, but I would frequently get a "Now the user wants me to switch roles..." Types of thinking. So I used deepseek-chat to clean up the reasoning (Which is not included here.) |
|
|
|
|
|
Lastly, I reprocess the conversation, removing my "user" instruction. This left ALL of the conversation as the "assistant" role. So I iterated through this and alternated between user and assistant. |
|
|
|
|
|
All of these conversations were generated using thinking. I will also be uploading a "thinking" dataset as well later. |
|
|
|
|
|
Ultimately this has about 1300 conversations, all of which are 5-15 turns! I might try to add more later that are longer (I.E 15-30+ turns). |
|
|
|
|
|
I did have a rule I added to all of the system prompts though. I mandated that all characters are 18 years old, or older. |
|
|
|
|
|
*UPDATE - 6/10/25* |
|
|
|
|
|
I removed a lot of entries after someone on reddit pointed out that there were some overused cliches. I.E 'Shiver down spine', 'barely above a whisper'. I removed a lot of the big offenders and regenerated a good bit of the content. |
|
|
|
|
|
Someone also mentioned that the models did not do well when trying to keep a secret. So I added in a scenario that involves a user keeping a secret, even under devastating consequences. So we will see how that goes! |
|
|
|
|
|
|
