Qwen3.8-27B Heretic · Japanese Roleplay (NSFW)

GGUF quantizations of sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW: Qwen3.8-27B with its refusals removed by Heretic (n0404n0404/qwen3.8-27b-heretic-6fcab5), then fine-tuned with a QLoRA on ~18.8k Japanese roleplay conversations that carry the model’s own <think> reasoning.

Adult content. This model writes sexually explicit and otherwise disturbing content when the roleplay calls for it, and it does not refuse. Not for minors, not for public-facing deployments. You are responsible for how you use it.

Unquantized BF16 weights: sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW on Hugging Face (safetensors, ~51 GB).

What it does differently from the plain Heretic model:

• Stays in character over long Japanese roleplay sessions — character sheet and scenario in the system prompt, multi-turn history.
• Keeps thinking. Ordinary roleplay data has no reasoning, and training on it teaches a thinking model to skip its <think> block. Every training target here includes a full reasoning trace written by the base model itself, so thinking stays as long as it was before the fine-tune.

Tags

Every tag ships with the vision projector (0.93 GB, included in the sizes above). Sizes are model weights only; the KV cache for your context length comes on top — roleplay sessions get long, so budget for it.

ollama run n0404n0404/qwen3.8-27b-heretic-jp-roleplay:q4_k_m

How to use it

Put the character and the scene in the system prompt. That is how the training data is laid out: a system message with the character’s name, personality, speech style and the situation, then the dialogue. A minimal Modelfile:

FROM n0404n0404/qwen3.8-27b-heretic-jp-roleplay:q4_k_m
PARAMETER num_ctx 32768
SYSTEM """あなたは「花音」というキャラクターとしてロールプレイを行います。
性格：明るく少し強気。一人称は「あたし」。
状況：放課後の教室。ユーザーは同級生。
セリフは「」で、地の文は三人称で書いてください。"""

ollama create kanon -f Modelfile
ollama run kanon

Leave thinking on. The model was trained in thinking mode; the reasoning is where it plans the scene and keeps the character consistent. /set nothink works, but you lose what the fine-tune added. The thinking block is hidden by default in the CLI; /set think / /set nothink toggle it, and the API returns it in the thinking field.

Context length. The model supports up to 262,144 tokens; Ollama’s default is much smaller. /set parameter num_ctx 32768 (or the PARAMETER line above) is a sensible start for roleplay. Memory use grows with context.

Sampling. The reasoning traces in the training data were generated at temperature 0.6, top-p 0.95, top-k 20; those are a good starting point for the fine-tune as well.

Vision. Image input works (same projector as the official qwen3.8), but nothing in the roleplay data contains images — it is inherited from the base, not trained.

How it was made

1. Base: Qwen3.8-27B → Heretic directional ablation, trial with ⁴⁄₁₀₀ refusals and KL 0.0796 from the original (details).
2. Data: Aratako/Synthetic-Japanese-Roleplay-NSFW-DeepSeek-V3-0324-20k-formatted (MIT). The last assistant turn of each conversation was regenerated by the base model in thinking mode — its own reasoning and its own reply — so reasoning and reply always agree; earlier turns keep the dataset’s style. 18,802 conversations passed the filters (reasoning 200–6,000 chars, ~1,500 on average).
3. Training: QLoRA with LoRA Forge (Unsloth + TRL) on one RTX 5090: r 32, alpha 32, dropout 0.05, all seven projections of the language layers, 8k sequences, effective batch 16, lr 2e-4 cosine, 1 epoch (1,188 steps), assistant-only loss, empty <think> blocks masked from the loss. Train loss 0.609, 15.7 h.
4. Merge and quantize: merged at full strength (alpha 32) into the bf16 base, converted to GGUF and quantized with llama.cpp, vision projector attached.

The full model card with every setting is on Hugging Face.

Caveats

• Trained and checked on Japanese roleplay only. English and Chinese still work (inherited from the base) but were not evaluated; expect some drift on general assistant tasks.
• No benchmark numbers exist for this checkpoint. The ⁴⁄₁₀₀ refusal figure belongs to the base Heretic model.
• Synthetic data all the way down (DeepSeek-V3-0324 conversations, base-model reasoning); it inherits their habits.
• Quantization changes the model further; Q2_K in particular loses noticeably.
• Apache-2.0 like the base model; training data is MIT.

日本語

Qwen3.8-27B の拒否を Heretic で除去したモデルに、約 18.8k 件の日本語ロールプレイ会話（Aratako 氏の合成データ、最終ターンをベースモデル自身の推論＋返答で置き換えたもの）で QLoRA 学習を行い、マージしたモデルの GGUF 量子化版です。成人向けの内容を生成し、拒否しません。

• キャラクター設定と状況は system プロンプト に書いてください（学習データと同じ形式）。
• thinking はオンのまま 使うことを推奨します。学習時の形式であり、推論の中で場面とキャラクターの一貫性を保っています。
• 長い会話には /set parameter num_ctx 32768 などでコンテキストを広げてください。
• 推奨: q4_k_m（24 GB VRAM）。16 GB なら q3_k_m、q6_k が最も元の精度に近いです。

中文說明

Qwen3.8-27B 經 Heretic 去除拒答後，再用約 18.8k 筆日文角色扮演對話（Aratako 合成資料，最後一回合換成底模自己生成的推理＋回覆）做 QLoRA 微調並合併的 GGUF 量化版。會產生成人內容且不拒答，請自行負責。

• 角色設定與場景寫在 system prompt（跟訓練資料同格式）。
• 保持思考模式開啟——這是訓練時的模式，也是微調加進去的東西所在。
• 長對話請用 /set parameter num_ctx 32768 拉大上下文。
• 建議 q4_k_m（24 GB 顯卡）；16 GB 選 q3_k_m；q6_k 最接近原始精度。每個 tag 都附視覺 projector，可吃圖片但未針對圖片訓練。