hotdogs/qwen3.6-27b-fable5-lora

Qwen3.6-27B LoRA Adapters

LoRA adapters for Qwen3.6-27B — fine-tuned for reasoning, function calling, and cybersecurity.

Latest Adapters

Adapter	Base	Description	Rank
qwen27b_opus4.8_sft 🔥	Huihui-Qwen3.6-27B-abliterated	Opus 4.8 reasoning SFT (6,956 rows)	r=64
VER4-SFT-ckpt-1200	Huihui-Qwen3.6-27B-abliterated	Fable perfect-v1 SFT (3,376 rows)	r=64
VER4-ORPO-v4.1-ckpt-95	Fable-ver4	ORPO fix: null/rambling/format	r=64

LoRA GGUFs (for llama.cpp)

File	Description	Size
GGUF/ver4-sft-lora.gguf	Fable SFT	638 MB
GGUF/ver4-orpo-v4.1-lora.gguf	ORPO v4.1	638 MB

Usage

Python (Transformers)

from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "huihui-ai/Huihui-Qwen3.6-27B-abliterated",
    torch_dtype="bfloat16"
)

# Load Opus 4.8 reasoning LoRA
model = PeftModel.from_pretrained(
    model,
    "hotdogs/qwen3.6-27b-fable5-lora",
    subfolder="qwen27b_opus4.8_sft"
)

llama.cpp

./llama-cli -m base_model.gguf --lora-scaled ver4-sft-lora.gguf:0.4

Training Details

Training	Data	Config	Time
Opus 4.8 SFT 🔥	6,956 rows	seqlen=8192, 1 epoch	~8h
Fable SFT (ver4)	3,376 rows	seqlen=4096, 3 epochs	~3h
ORPO v4.1	150 pairs	r=64, lr=5e-6, 5 epochs	~45m
Cyber SFT v2	70K filtered	r=16, seqlen=4096	~3h

---

Built with ❤️ by UKA — 18yo coder & cybersecurity expert

💖 Support / โปรดสนับสนุน

If you find these adapters useful, please consider supporting my work!
หากคุณคิดว่า LoRA นี้มีประโยชน์ กรุณาสนับสนุนผลงานของฉันด้วยนะคะ! 🙏

₿ Bitcoin — BTC:

bc1qf27cyk3vmugcdyv9xdtuv5jwz37863crpj5c9v

Thank you for your support! 🙏✨
ขอบคุณมากๆ สำหรับการสนับสนุนค่า! 💖🤗