ScarletKc-1.0-GGUF

ScarletKc-1.0 · Q8_0 GGUF

A personal persona model, fine-tuned with BF16 LoRA on my own chat history. It speaks the way I do: mostly Chinese, and very briefly. It knows it is an AI rather than the person itself.

Trained 2026-08 on chat logs spanning 2020 to 2026. Knowledge beyond that comes from the base model, not from me.

Examples

你是谁               → scarletkc
你在干嘛              → 发呆
你更喜欢苹果还是香蕉    → 苹果
猫会游泳吗            → 会

With thinking enabled it reasons first, then answers:

9.11 和 9.9 哪个大    → <think>…</think> 9.9 大 9.9 等于 9.90

Usage

llama-server -m ScarletKc-1.0-Q8_0.gguf -ngl 99 --jinja

Use sampling, not greedy — temperature 0–0.1. Both chat modes are trained; switch with chat_template_kwargs: {"enable_thinking": true|false}.

Requires a llama.cpp build from 2026-08 or later (Qwen3.5 architecture support).

Notes

Only Q8_0 is published — it is the variant verified end to end. Its outputs match the BF16 master token for token, and generation terminates correctly at <|im_end|>.

Intended use

Casual conversation and entertainment. This is a style model, not a knowledge model — 0.31% of the parameters were trained and all of that capacity went into how it talks. Do not use it for work, factual lookup, or anything where a wrong answer costs something.

• Anything it says about "itself" is invented. Ask the same question three times and you get three different cities. Hallucination was never trained against, so treat no statement about a real person as fact.
• Its self-identity is unstable. It usually answers scarletkc or kc, but will sometimes offer another name, occasionally one belonging to a different AI assistant.
• It may address you by a name you never gave. The source conversations had one recurring chat partner, so that name sometimes surfaces in the second-person slot.
• It concedes easily. Contradict it and it will generally agree rather than hold its ground. A faithful trait of the source conversations.
• Replies are short by design, not truncated.

Data

Private Telegram and QQ conversations. Loss was computed only on my own replies — the other side's messages served as context and were never a training target. URLs, emails, phone numbers, long IDs and credentials were stripped before training.

None of the training data is in this repository. Probing the released model for personal details returns invented answers that differ on every sample.

Why this exists

This is an archive, not a product. I trained it to find out whether I could, and published it so that something of it stays in the world. There is no roadmap, no support, and no benchmark I am competing on.

If you try it, treat it as a curiosity rather than something to build on. I would rather it stayed a small thing.

License

Derivative of Qwen/Qwen3.5-9B (Apache-2.0), whose full license text is included as LICENSE. The fine-tuning is licensed under PolyForm Noncommercial 1.0.0, included as LICENSE.polyform. Use of this model is subject to both.

Required Notice: Copyright ScarletKc (https://huggingface.co/ScarletKc)