base_model:
🔗 Base Model
This model is built on huihui-ai/Huihui-Qwen3.6-27B-abliterated — the abliterated (uncensored) version of Qwen3.6-27B with all refusal mechanisms removed.
Last updated: 2026-07-08
📦 Available Models
| Folder | Description | Download |
|---|---|---|
qwen27b-abliterated-Fable-opus4.8/ | Opus 4.8 reasoning SFT | ✅ |
Qwen3.6-27B-abliterated-Fable-opus4.8-cyber/ | Opus 4.8 + Cyber knowledge | ✅ |
| File | Format | Size | Description |
|---|---|---|---|
GGUF/...opus4.8-cyber.Q4_K_M.MTP.gguf | Q4_K_M | 16 GB | Recommended |
GGUF/...opus4.8-cyber.F16.MTP.gguf | F16 | 54 GB | Full precision |
GGUF/...opus4.8_F16_MTP.gguf | F16 | 54 GB | Base (no cyber) |
GGUF/...opus4.8_Q4_K_M_MTP.gguf | Q4_K_M | 16 GB | Base (no cyber) |
Qwen3.6-27B (base)
│
├── Huihui-Qwen3.6-27B-abliterated (diff-in-means)
│
├── ORPO v1.2 → v1.10 (10 iterations)
│ └── Best: v1.9 (margin 0.35)
│
└── VER4 (Clean Slate)
├── SFT perfect-v1 (3,376 rows, Hermes format)
├── ORPO v4.1 (150 pairs)
├── Cyber SFT v2 (70K filtered)
└── 🏆 Opus 4.8 SFT (6,956 rows, Claude-style reasoning)
📚 Dataset
| Dataset | Rows | Source | Description |
|---|---|---|---|
| perfect-v1 | 3,376 | Fable-5 | Reasoning + Hermes format conversion |
| perfect-v2 | 5,376 | Fable-5 + hermes-fc | Tool calling + reasoning |
| fable-opus-reasoning 🔥 | 6,956 | Fable-5 + Opus 4.8 | Pure reasoning, no tools |
🚀 Usage
llama.cpp (Recommended)
./llama-cli -m GGUF/Qwen3.6-27B-abliterated-Fable-opus4.8-cyber.Q4_K_M.MTP.gguf \
--temp 0.6 --top-k 25 --top-p 0.9 --min-p 0.1 \
--repeat-penalty 1.15 --dry-multiplier 0 --dry-sequence-breaker none \
--ctx-size 65536
Python
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"hotdogs/qwen27b-abliterated-Fable-MTP",
subfolder="Qwen3.6-27B-abliterated-Fable-opus4.8-cyber"
)
📖 References
| Paper / Project | Citation |
|---|---|
| LoRA | Hu et al. "LoRA: Low-Rank Adaptation of Large Language Models" (ICLR 2022) |
| ORPO | Hong et al. "ORPO: Monolithic Preference Optimization without Reference Model" (2024) |
| Hermes FC | NousResearch. "Hermes Function Calling" |
| Abliteration | Ardila et al. "Refusal in LLMs is mediated by a single direction" (2024) |
| Fable-5 | Glint Research. Multi-step reasoning agent traces |
| Qwen3.6 | Qwen Team. "Qwen3.6: Scaling Open Language Models" |
🐛 Known Fixes
| Issue | Fix |
|---|---|
Model stops at colon (:) | --dry-multiplier 0 --dry-sequence-breaker none |
| Infinite correction loops | Fixed in opus4.8-cyber release |
| MTP tensor missing | All GGUFs now verified (866 tensors) |
💖 Support / โปรดสนับสนุน
If you find this model useful, please consider supporting my work!
หากคุณคิดว่าโมเดลนี้มีประโยชน์ กรุณาสนับสนุนผลงานของฉันด้วยนะคะ! 🙏
₿ Bitcoin — BTC:
bc1qf27cyk3vmugcdyv9xdtuv5jwz37863crpj5c9v
Thank you for your support! 🙏✨
ขอบคุณมากๆ สำหรับการสนับสนุนค่า! 💖🤗