Qwen3.5-9B-Melancholy-Viber-GGUF
A QLoRA fine-tune of Qwen/Qwen3.5-9B that replies to short, emotionally-loaded remarks the way people do in the comment section of NetEase Cloud Music — terse, sincere, a little melancholic, occasionally devastatingly witty.
Provided as a single Q4_K_M GGUF (5.3 GB, 5.02 BPW) for llama.cpp / Ollama / LM Studio.
Real outputs from this model
input: 28岁了，一事无成 output: 28岁，才刚刚开始。
input: 风把云吹散了 output: 是啊，有时候，想通了，就散了。
input: 养了十二年的狗今天走了，它最后是看着我出门的 output: 它不会说话，但它用行动教会了我什么是忠诚与陪伴。
Training
- Method: QLoRA (nf4, double-quant), r=32 / alpha=64, targeted at all attention + MLP projections, 1 epoch, lr 2e-5 (cosine), effective batch 16.
- Data: 4,312 comment → reply pairs mined from ~200k NetEase Cloud Music comments (top-100 hot comments per song across 2,000 pure-music tracks; only replies endorsed by 200+ likes). Replies are verbatim human replies — never generated.
- Thinking: each reply is anchored by a short human-like inner monologue (
<think>channel, trained atreasoning_effort=low). The monologue is machine-distilled, the reply is real.
Usage
llama-server -m Qwen3.5-9B-Melancholy-Viber-Q4_K_M.gguf --jinja --temp 0.4 --repeat-penalty 1.1
Recommended: temperature <= 0.5, repeat_penalty >= 1.1. The reply is preceded by a short <think> inner monologue (shown as reasoning in most UIs).
Known limitations
- Ultra-short poetic inputs (≤7 chars) can still trigger echo or repetition — mitigate with
repeat_penalty 1.15and resampling. - Numeric/math reasoning is fragile after emotional-domain fine-tuning.
- Replies intentionally stay terse and unadvice-giving; it will not comfort you the way a general assistant does. That is the point.
Intended use
Research and personal use. The training data comes from public comment sections; replies are verbatim user-generated content — treat outputs accordingly.