Differences Between SillyTarvern and Third-Party Platforms

At present, I generally avoid using overly complex character cards that rely on JSON code. While most large language models can read and execute JSON—evident from the sophisticated JB presets shared in the community—many third-party platforms outside the Tavern (such as Chub) often fail to support this due to platform limitations and model constraints. The model’s training data, context window size, and integrated functionalities often lead to JSON being rejected or programming capabilities being lost. Moreover, output truncation due to token limits is a common issue.

Known Limitations on Chub

• Free/Mobile Models
  • Multilingual Support: Almost nonexistent. While these models can understand and output basic text, non-English corpora are extremely limited. The content seems to stem from the base model’s pre-finetuning data (possibly Qwen?).
  • Chain-of-Thought Reasoning: Not supported at all. These models neither interpret nor generate reasoning chains.
  • Formatted Language: Minimal support. Chub can render inline HTML formatting, and free models can generate Markdown, but formatting is often lost due to output limitations.
  • Coding Capabilities: Not supported. These models cannot read JSON and have poor comprehension of HTML structures.
  • Context Size: Very small. They struggle with complex LoreBooks and sometimes even fail to interpret the content of the character cards themselves.
  • Summary: Essentially just sweet spot. Not recommended for any serious use.
• Tier 1 Paid Models – Mercruy Rec & Mercruy
  • Multilingual Support: Very limited. They can parse and output text, but rarely generate anything outside alphabetic languages.
  • Chain-of-Thought Reasoning: Completely unsupported.
  • Formatted Language: Mostly supported. These models can output Markdown and HTML reliably, though still face occasional issues due to context limitations.
  • Coding Capabilities: Still unsupported. JSON reading is unavailable, and HTML handling is average at best.
  • Context Size: Reasonable. Most character cards function properly. The 13B Rec variant performs particularly well.
  • Summary: Usable, but limited in functionality.
• Tier 2 Paid Models – Asha / Full Power Mercruy Rec
  • Haven’t tested personally, so I can't comment. However, if Asha really has 80B parameters, it should function like most full-scale LLMs.
• Third-Party Platform Models
  • Currently, Chub supports third-party platforms such as: all models from Anthropic (Claude series), selected models from OpenAI (pre-GPT-4o), early versions of Google's models (prior to Gemini Pro), some models from NovelAI (untested by me), all endpoints on OpenRouter (any API-supported model), and self-hosted open-source models via Kobold.
  • I highly recommend Anthropic’s Claude and OpenAI’s ChatGPT. I do *not* recommend Google's early Gemini models due to their outdated performance. OpenRouter is a great third-party option, and users can try out classic open-source models like Qwen, various Hugging Face releases, and DeepSeek. For Chinese character cards, DeepSeek R1 and V3 are likely the best options available. Gamma, fine-tuned on Claude’s corpus, and the Tifa series trained on DeepSeek R1, are also excellent choices.
  • While the API reviews for ChatGPT and Claude are less strict than their official web platforms, models from GPT-4o onwards have heavy moral restrictions. Claude 3.7 Sonnet currently does not enforce strict moderation, but both OpenAI and Anthropic APIs have been known to implement post-hoc content audits, which can result in account bans. Considering the cost, I don't recommend using ChatGPT or Claude for NSFW roleplaying purposes.

As for the configuration of Chub as a chat platform itself, I’m still not entirely sure. It uses its own tokenizer, which doesn't count most tokens the same way as other providers do. This leads to unusual behavior when handling non-Latin languages, often triggering truncation on the platform. Overall, I only recommend using English with models on Chub; most other languages won’t run reliably, and you should avoid using complex LoreBooks or code-heavy structures altogether.

The Template I'm Currently Using

The structure of my character cards is mostly based on templates recommended by ST. While the LoreBook section can technically be filled freely, Chub imposes a strict 8196-token limit. Also, some features within the LoreBook may break when imported into the latest version of ST. Although it usually still works, I don’t recommend using overly complex structures in the LoreBook.

A good character card should keep the key information—i.e., the part of the input the LLM pays the most attention to, the Description—under 1500 tokens, ideally around 1000. Using English helps conserve tokens and improves memory retention and accuracy. It’s okay to keep a few terms (e.g., names, places) in Chinese, or even write the whole thing in Chinese, but it’s not recommended. It wastes tokens and increases comprehension difficulty for the model. From a linguistic perspective, different languages are not always directly translatable, and meaning can drift—an LLM may gradually lose track of non-English content as more history accumulates.

The format is based on JSON-style structure. The syntactic symbols of JSON (quotes, braces, brackets) don’t usually count as tokens, and a structured format helps LLMs parse content more effectively. It doesn’t need to be strict JSON; the idea is just to make your card look like structured data, which aligns better with how ST and LLMs process input.

Template Structure

Section 1: Introduction block. Although content in the Scenario field gets injected into the input, I usually integrate important Scenario details into the higher-priority Description block due to LLM attention issues.

This section explains the LLM’s job and the world setting. First, I define the main setting—e.g., modern nation, medieval fantasy, Japanese light novel, horror thriller, etc. Then, I add guardrails to prevent the LLM from using out-of-style vocabulary and to strengthen the tone. Finally, I clarify the LLM’s role (e.g., narrator) and prevent it from roleplaying as the user. I also include language and stylistic notes here—since most LLMs are trained with popular literature, mentioning series like Harry Potter, A Song of Ice and Fire, The Lord of the Rings, or even One Hundred Years of Solitude helps. Light novels like Re:Zero or KonoSuba also seem to appear frequently in recent models, likely due to wide fan translation circulation.

Section 2: World and character summary. I usually add a twist in character design. For example, in Salutte, I first state that Salutte assassinated the king and fled, but later reveal it was due to a noble threatening her lover. This section is split into Overview (broad context and tone) and Fact (hidden story or plot hook).

Section 3: Character description. While I label this section as Appearance, it includes Age/Weight/Height/Hair/Eyes/Face/Skin/Body/Features. “Body” refers to general appearance and first impressions, more physical traits, or fantasy elements like cat ears or scars. “Features” describes personality and everyday traits. I usually include clothing, carried items, and scent.

Section 4: Abilities. This expands on powers, skills, or unique traits—whether magical, combat-related, or hobbies. It helps deepen your character and gives you material for plot development.

Section 5: Personality. This is crucial. It determines how the LLM will portray your character. I strongly advise against using extreme descriptors in tags like “angry,” “fragile,” “sensitive,” or “insecure.” Even if that’s your goal, use neutral or softened language. While ChatGPT and Claude are safeguarded by ethical filters, most LLMs become hysterical with extreme traits. Use Like/Dislike fields to indicate abstract social preferences like “being praised” or “feeling isolated”—these help spark ideas.

Section 6: Lifestyle and habits. This section builds external context—items they carry, where they live, goals, daily routines. It’s broad and meant for your own inspiration. Avoid overly complex life goals here; keep those in Overview/Fact or the LoreBook.

Section 7: Orientation. Ah yes, the NSFW bit. If your character has... special preferences, define them here. I usually default to pansexual to accommodate different user identities—I design Anypov characters, after all. As for the Kinks section... well, feel free to unleash your creativity.

Example Style

You can refer to the cards I’ve published on Chub.

Other Notes

Scenario: This functions like the introduction. Use a few keywords to describe the desired tone and list what the LLM shouldn’t do. This is optional though, as most modern LLMs will understand you well enough.

Example Dialogs: These train the LLM to imitate the character’s speaking style. Use ST’s official format for reference. This section is important—list your character’s behavior in various moods (calm, happy, angry) and situations (praying, training, working, school), and of course... NSFW scenarios.

LoreBook Design: I’ll cover that in another post. It’s getting late...