GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
sss22213
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
sss22213
Model Tree
sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW
Input Modalities
Output Modalities
Supported Functionality
| Step | What | Where |
|---|---|---|
| 1 | Qwen3.8-27B (Apache-2.0, vision-language, thinking) | Qwen/Qwen3.8-27B |
| 2 | Refusal direction ablated with Heretic (trial 198: 4/100 refusals, KL 0.0796) | sss22213/Qwen3.8-27B-Heretic-NoRefusal |
| 3 | Japanese roleplay QLoRA on reasoning-augmented data, merged at full strength | sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW |
| 4 | Scene → Danbooru tags LoRA trained on top of step 3, merged at quarter strength | this repository |
blind-captionsSource: dartags/danbooru-2408-blind-captions (MIT, 6.33M rows). Each row pairs a one-sentence Japanese description of a Danbooru post (summary_ja, written by an LLM from the post's tags, no image involved) with the post's tags. It is the dataset behind DanbotNL.
The rows were converted into chat samples (system / user / assistant) with LoRA Forge's prepare_blind_captions.py:
summary_tags, not general. summary_tags are the tags the description actually mentions; general is every tag on the post, including things invisible in the text (art style, year, censored). Training on general would teach the model to invent tags, which is the opposite of the goal.artoria pendragon \(fate\)), tags are comma-separated on one line in the order character → copyright → described tags. Posts rated questionable/explicit get a leading nsfw.Final training set: 450,821 samples (440,821 scene → tags, 10,000 roleplay). The system prompt used for every tag sample (use the same one at inference):
あなたはイラスト生成用プロンプトの専門家です。与えられた場面描写を、Stable Diffusion WebUI（A1111）向けの Danbooru タグに変換してください。描写に含まれる要素だけをタグにし、存在しないタグを作らず、タグはカンマ区切りで一行に出力してください。
Trained with LoRA Forge (Unsloth + TRL SFT) on a single H100 SXM 80 GB (RunPod), in bf16, on top of the merged Qwen3.8-27B-Heretic-JP-Roleplay-NSFW weights rather than the original base, so the adapter learns the change relative to the roleplay model it will be merged into.
| Setting | Value |
|---|---|
| Base weights | Qwen3.8-27B-Heretic-JP-Roleplay-NSFW, bf16 (no quantization) |
| LoRA | r = 32, alpha = 32, dropout 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj — language-model layers only |
| Sequence length | 8,192 (no packing) |
The adapter was merged into the bf16 Qwen3.8-27B-Heretic-JP-Roleplay-NSFW weights with PEFT merge_and_unload at lora_alpha = 8 (effective scale alpha / r = 0.25). At the trained strength (alpha 32) and at alpha 16 the merged model's reasoning collapsed and its output degraded; alpha 8 keeps the roleplay behaviour and reasoning length of the parent while retaining the tag vocabulary. The learning rate of this run (4e-4) was four times the intended value, which is why a quarter-strength merge lands where a full-strength merge of a 1e-4 run would. Output: two safetensors shards, ~51 GB, Qwen3_5ForConditionalGeneration, unquantized.
Use it exactly like Qwen/Qwen3.8-27B — same chat template, same enable_thinking / reasoning_effort controls.
enable_thinking=False, or /no_think). The tag samples were trained with an empty think block in context, so this matches training exactly and skips a reasoning trace you do not need. The reply is a single comma-separated line ready to paste into an A1111 prompt box (add your own quality tags and negative prompt).python
from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "sss22213/Qwen3.8-27B-Heretic-JP-Roleplay-NSFW-DanbooruTags"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto") SYSTEM = ("あなたはイラスト生成用プロンプトの専門家です。与えられた場面描写を、" "Stable Diffusion WebUI（A1111）向けの Danbooru タグに変換してください。" "描写に含まれる要素だけをタグにし、存在しないタグを作らず、タグはカンマ区切りで一行に出力してください。")messages = [ {"role": "system", "content": SYSTEM}, {"role": "user", "content": "電車の中で、黒髪の女子高生が窓際に立ち、頬を赤らめながらこちらを見ている。"},]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", enable_thinking=False,).to(model.device)output = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))# e.g. 1girl, solo, black hair, school uniform, train interior, standing, window, blush, looking at viewerThe full checkpoint is ~51 GB; for a single consumer GPU, quantize it (GGUF via llama.cpp / Ollama). Because the vision encoder is the stock Qwen3.8-27B one, a GGUF import can reuse the official Qwen3.8 vision projector for image input.
blind-captions are one sentence about a single picture. Converting a long multi-paragraph roleplay scene works, but the model was not trained on that input shape; shorter, picture-like descriptions give the most precise tags.summary_tags. The source captions and tag subsets were generated by LLMs from Danbooru metadata; artist tags, meta tags and anything not stated in the description are intentionally never produced.danbooru-2408-blind-captions and the DanbotNL work it comes from.Released under Apache-2.0, like the base model; the training data is MIT.
lolishotaaged downaratako-rp-think1, with their full <think> traces) were shuffled in so the run keeps seeing long reasoning while learning short tag answers.| Batch | 4 × gradient accumulation 4 = 16 sequences |
| Optimizer / LR | AdamW 8-bit, 4e-4, cosine schedule, 400 warm-up steps, weight decay 0.01 |
| Epochs | 1.01 (28,459 steps), seed 3407 |
| Loss | assistant turns only |
Empty <think> handling | mask — tag samples carry no reasoning; their empty think blocks are excluded from the loss so the run never trains the model to skip thinking |
| Result | train loss 0.507, 33.1 h, peak VRAM 61.9 GB |