Base Model: llmfan46/Qwen3.5-27B-ultra-uncensored-heretic-v2
Architecture: Qwen3.5-27B (Dense, 27B parameters, all active)
Training Method: LoRA fine-tuning with Unsloth, merged to bf16
Training Data: Jongsim/claude-opus-4.6-reasoning-12k-en-filtered-v2 (12,822 examples)
Format: SafeTensors (bf16)
Size: ~54.7 GB

Training Configuration

Parameter	Value
LoRA rank (r)	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Epochs	3
Batch size	1 (GA=8, effective=8)
Learning rate	2e-4
Scheduler	Cosine
Max sequence length	2048
Optimizer	AdamW 8-bit
Precision	bfloat16
Hardware	NVIDIA DGX Spark (GB10 Blackwell GPU, 128GB unified memory)
Total training time	~113 hours

Training Loss

Epoch	Start Loss	Final Loss	Avg Loss	Improvement
1	0.5551	0.2747	0.3375	—
2	0.2738	0.1283	0.1757	-47.9%
3	0.1287	0.0585	0.0735	-58.2%

Overall training loss: 0.1954

The model shows strong convergence with consistent loss reduction across all 3 epochs. No signs of overfitting observed.

Dataset

The training dataset consists of 12,822 high-quality English reasoning examples generated by Claude Opus 4.6, featuring:

Complex multi-step reasoning with chain-of-thought
<think>...</think> structured reasoning traces
Diverse domains: math, logic, coding, analysis, creative writing
Quality-filtered for coherent and complete reasoning chains

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "Jongsim/Qwen3.5-27B-heretic-v2-Opus-4.6-Distilled"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Explain the concept of gradient descent in machine learning, step by step."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Attribution

Base model abliteration: llmfan46 — created the uncensored heretic variant
Original architecture: Qwen Team — Qwen3.5-27B
Fine-tuning & distillation: Jongsim — LoRA training with Claude Opus 4.6 reasoning data
Training framework: Unsloth

License

This model inherits the Apache 2.0 license from the base Qwen3.5 model.