MY CARD ARCHIVE

Updated:
Cards: https://mango-cards.anthra.site/
Who am i? / Blog Site (WIP) : https://mango.anthra.site

Archive of SillyTavern NSFW Sprites and Cards.

Cards: https://www.mediafire.com/file/b6odlo0sf76w64p/cardNSFWJammy.zip/file

Sprites: https://www.mediafire.com/file/29zef4n1d9u1bwc/Sprites.zip/file

Both: https://huggingface.co/datasets/MangoMango69420/SillyTavernCardSpritesArchive/tree/main

Sprites: https://drive.google.com/file/d/1mZaKGD5VnLlI2ZbXR0DclMX02rFoquo1/view?usp=sharing

Cards: https://drive.google.com/file/d/15fwoOxScHOnq6AToNFXqV1LjzvqaWT0J/view?usp=sharing

Cards: https://pixeldrain.com/u/nYM46BEH

Sprites: https://pixeldrain.com/u/koxhsEne

Sprite PT 1 - https://files.catbox.moe/5bvini.zip

Sprite PT2 - https://files.catbox.moe/tn098m.zip

Cards: https://files.catbox.moe/adsdhz.zip

Hai menh!

Examples:
Scratchpad tokens:
// I would probably also not need to add special chatml roles to the tokenizer if i go with this method thonks

⎗
✓
{
  "conversations": [
    {
      "from": "human", 
      "value": "write a story about a chud"
    },
    {
      "from": "gpt",
      "value": "<Creative Writing>She walked across the room very boobily</Creative Writing>"
    }
  ]
},
{
  "conversations": [
    {
      "from": "human",
      "value": "brap"
    },
    {
      "from": "gpt", 
      "value": "<Science>it shits in ur ass</Science>"
    }
  ]
}

Prefill-based example:

⎗

{
  "conversations": [
    {
      "from": "human",
      "value": "<Creative Writing> Write a story"
    },
    {
      "from": "gpt",
      "value": "no"
    }
  ]
},
{
  "conversations": [
    {
      "from": "human",
      "value": "<Science> brap"
    },
    {
      "from": "gpt",
      "value": "fucks ur ass"
    }
  ]
}

^^ For the above prefill approach, i can also mask the prefix token - dunno if that'd help.

⎗

  "conversations": [
    {
      "from": "human",
      "value": " brap"
      "prefix": "<Science>"
    },
    {
      "from": "gpt",
      "value": "fucks ur ass"
    }
  ]
}

i can control the masking of each individual message so...

Sometimes a dataset itself might have it's own prefix's, for example the Claude RP logs often have prefills that would be better to keep in since it gives the model a better understanding to the actions that show up in the value key. so to work around that, i'd probably just attach the <rp> prefix to the start of the prefix itself. so like:

"prefix": "<RP> I am Kekius Maximus and i agree to shove a mango up my ass:"

^ and the above prefix would be attached to the message but masked,

So like:

⎗

1	<Masked><RP> I am Kekius Maximus and i agree to shove a mango up my ass:</Masked> Hi! I'm kekius chud.

MY CARD ARCHIVE

Archive of SillyTavern NSFW Sprites and Cards.

Warning