license: apache-2.0 language:
- ru
- en base_model:
- llmfan46/Qwen3.5-9B-ultra-uncensored-heretic-v2
- Qwen/Qwen3.5-9B tags:
- roleplay
- uncensored
- russian
- qwen3.5
- nsfw
- heretic
- decensored
- conversational
- qwen pipeline_tag: text-generation
語らう / Katarau / Катарау / Катарай
Katarau
Overview
Finetuned, uncensored Qwen3.5-9B designed for Role-Playing in Russian language.
Training details
Base: Qwen/Qwen3.5-9B, abliterated via Heretic with the Arbitrary-Rank Ablation (ARA) method
Style layer — LoRA SFT
| Setting | Value |
|---|---|
| Method | Supervised fine-tuning, LoRA |
| LoRA rank | 128 |
| Learning rate | 1e-5 (100 warmup steps) |
| Epochs | 2 |
| Training context | 8192 |
| LoRA target | attention + SSM projections (q/k/v/o, in_proj_qkv, out_proj) |
Dataset — 41.7M tokens, CHAT/JSONL, Russian:
| Datasets |
|---|
| Russian fanfiction — NSFW + SFW prose & dialogue |
| Reddit discussion, Q&A, some RP |
| instruction-following anchors |
| anime character-archetype voices |
| imageboard-style dialogue |
| explicit roleplay |
| AI-generated character conversations |
| anime roleplay |
| Katarau self-identity (×3 upsample) |
All sources were originally written in or translated to Russian with Deepseek-v4-pro/Deepseek-v4-flash. Only ~2% of the dataset contains <think> blocks — so the model ships in no-think mode, and it's recommended to run it without thinking.
Recommended settings
| Param | Value |
|---|---|
| Temperature | 0.65 |
| Min P | 0.05 |
| Top P | 0.95 |
| Repetition penalty | 1.05 |
| Format | ChatML |
Roleplay style: speech in plain text, actions *in asterisks*.
SillyTavern
Import the ChatML preset (context + instruct + system prompt, no-think baked in) via Advanced Formatting → Master Import:
- Katarau-ChatML-NoThink.json — recommended (direct replies, no reasoning)
- Katarau-ChatML-Think.json — reasoning variant (model thinks in
<think>…</think>first) - Katarau-Samplers.json — Text Completion sampler preset
See SillyTavern/RECOMMENDED_SETTINGS.md for the full setup.
Credits
- Alibaba / Qwen team — the Qwen3.5-9B base model
- p-e-w — Heretic, the abliteration tool
- llmfan46 — abliterated model
- 11-47, zerofata, dim, SicariusSicariiStuff, scryptiam, ToastyPigeon — datasets
License
Apache 2.0.