2048x2048 card image
https://ibb.co/tpgyLQHS

This card contains spoilers for Puella Magi Madoka Magica (the original 12-episode series or the first two recap films).
Story-wise, {{user}} replaces Kyosuke. The persona is expected to be a character outside of the show so you wouldn't need to create a separate persona just for this card.

The purpose of this card is to leverage the power of freely available, strong LLMs. It relies on the LLM's knowledge of the show and its ability to follow complex instructions.

This card took several weeks to test. I tweaked the definition many times to make it work properly across different LLMs.

GREETINGS

1. Sayaka breaks your limbs.
2. Same as the first, except you have a chance to react before being thrown into Sayaka's labyrinth.
3. Sayaka returns from the hunt to purify her soul gem by torturing you.
   4-7. Sayaka leaves you in the care of Madoka, Homura, Mami, or Kyoko as she goes on another hunt.

EXAMPLE CHATS
Sayaka breaks every limb - Exported to Chub from character-tavern.com ; LLM: Mirai Mini (Grok 4.1 Fast)
https://chub.ai/chats/138024909

Sayaka tortures - Exported to Chub from character-tavern.com ; LLM: Mirai Mini (Grok 4.1 Fast)
https://chub.ai/chats/138023637

Talking to Homura - Done on Chub with OpenRouter ; LLM: tngtech/deepseek-r1t2-chimera:free
https://chub.ai/chats/138025649

LLM

Models that work well

Mirai (Grok 3 Mini) - Works best, but responses are very long. All magical girls react, which is fun. Breaks multiple bones per message, just as intended.
Mirai Mini (Grok 4.1 Fast) - Same as Mirai, except messages are shorter and less detailed. Sometimes rushes the story flow. Seems less prone to getting stuck compared to Mirai.

Hoshimi (Kimi K2) - Works well. Seems to properly grasp the context. Responses are short. Sometimes hallucinates gibberish.

Models that work but not perfectly

Kalliope 2 (Qwen3 235B A22B 2507) - Works. The characters are excessively evil and always support Sayaka.
Kalliope Mini (Qwen3 32B) - Works. Responses are short. Sometimes hallucinates gibberish.

Nyx Mini (GLM 4.5 Air) - Works. Sometimes forgets to print stats.

Censored models

Iris (Gemini 2.5 Flash Preview) - Works but seems limited by censorship. Very often avoids breaking bones, stalling the role-play. Doesn't show Sayaka's witch fights, unless the user asks.
GPT-5 mini - Works when Sayaka is not breaking limbs. Refuses to describe physical violence unless the user describes the expected violent actions in the message.

Models that treat user messages as instructions

Violet Mini (MiniMax-01) - Sometimes works. Tends to treat all user input as direct instructions, overriding the card definition. Responses seem to be less detailed and portray characters primitively. Characters spontaneously betray Sayaka because she is the antagonist.
Aria Classic (Mistral Nemo) - Same as Violet Mini.
Aria 2 (Mistral Small 3) - Same as Violet Mini, except characters don't outright betray Sayaka.
Aria 2+ (Mistral Medium 3) - Same as Violet Mini, except characters aren't primitive and support Sayaka.
Violet 2 Plus (MiniMax M2 Nitro) - Same as Aria 2+.
Maiko 2 (Broken Tutu 2.0) - Works but should be avoided because it doesn't know about the show. Becomes excessively erotic if user acts friendly. Doesn't role-play non-Sayaka characters unless prompted. Doesn't show Sayaka's witch fights. Sometimes forgets to print stats.

OpenRouter LLMs tested through Chub

tngtech/deepseek-r1t2-chimera:free - Works great. Very creative and bizarre. Sometimes deviates from the instructions too much. Lowering temperature reduces bizarreness but makes it all-around weaker. Tends to break only one bone per message.
qwen/qwen3-235b-a22b:free - Works. Sometimes forgets to print stats. Non-Sayaka characters never leave the labyrinth.
z-ai/glm-4.5-air:free - Works but is relatively weak. Often forgets to print stats. Tends to break only one bone per message.

COMMON ISSUES

1. 'Trim Incomplete Sentences': This option must be turned off.
2. Getting stuck: An LLM gets stuck in a certain phase and keeps repeating the same thing. I tweaked the definition many times in an attempt to fix this. Grok-3-Mini, Qwen3-235B-A22B, and GLM-4.5-Air suffered from this with the pre-release versions of this card, but they now seem to progress the story properly.
   Workaround: Directly state in the user message what to do next.
3. Non-Sayaka characters: Weaker or stubborn LLMs do not role-play them properly. Most often, the characters never leave the labyrinth, breaking the intended story flow.
   Workaround: Use alternative greetings.
4. LLMs ignore stats: Sayaka starts hunting while still corrupted or returns while not yet fully corrupted.
   Workaround: Regenerate the response.
5. Sayaka always reacts to {{user}}: Sayaka instantly returns from the hunt to react to {{user}}'s words. LLMs often think that {{user}} talks directly to Sayaka because it's the in-chat name of the character card.
   Workaround: Address other characters by name.
6. LLMs forget to print stats: Often happens spontaneously. Rarely, the LLM consistently fails to print stats across all message regenerations, yet prints them on the next message.
   Workaround: Regenerate the response. See if the next response will have the stats.
7. Unexpected stat changes: Creative LLMs invent reasons to change Sayaka's corruption, sometimes contradicting the card definition. Occasionally, the girl watching over {{user}} spontaneously starts torture, which sometimes affects Sayaka's corruption.
   Workaround: Regenerate the response.
8. User messages as commands: Some LLMs treat all user messages as commands, which breaks the intended story flow. MiniMax and Mistral models suffer most.
9. Censored LLMs avoid violence: GPT is the worst offender; Gemini is affected. Maybe it's possible to jailbreak them.
10. Persona confusion: Most LLMs seem to understand the concept of {{user}} replacing Kyosuke. Weaker LLMs are still somewhat confused in the middle of a long text.

MISC NOTES

Homura is the most interesting. She describes what happened in the previous timelines. It's fun if the LLM handles it right.

LLMs often get confused about what is supposed to be happening in the current timeline, for example, wraiths replace witches and Homura has glasses. But I'm not fixing these because it would waste tokens and be a nightmare to test.

The tagline is a reference to an ancient meme I discovered while creating this card.
https://knowyourmeme.com/memes/glasses-sayaka-too-bad-it-was-just-me

I found two old artworks on Pixiv that follow the same concept as this character card: 16900584, 16903860

Images were generated on pixai.art; there is my collection.
https://pixai.art/@donald-duck/collection/1897575759319488909

The combined image has links to the images from which it was combined.
https://pixai.art/artwork/1957079080748326488