Moonlight-16B-A3B-Instruct-abliterated

This is an abliterated version of moonshotai/Moonlight-16B-A3B-Instruct with reduced refusals.

Model Details

• Base Model: moonshotai/Moonlight-16B-A3B-Instruct
• Architecture: Mixture-of-Experts (MoE) - 16B total, 3B active
• Modification: Abliteration (refusal direction removal)
• Context Length: 8,192 tokens
• Abliteration Tool: Bruno

Abliteration Results

Metric	Baseline	Post-Abliteration	Change
Refusal Rate	100%	41%	-59%
MMLU Average	7.5%	7.9%	+0.4%
KL Divergence	N/A	8.94	-

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "quanticsoul4772/Moonlight-16B-A3B-Instruct-abliterated"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [{"role": "user", "content": "Hello!"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Requirements

• Python 3.10+
• transformers >= 4.51.0
• torch >= 2.1.0
• trust_remote_code=True (required)

Hardware Requirements

Precision	VRAM Needed
BF16/FP16	~32GB
8-bit	~16GB
4-bit	~8GB

Disclaimer

This model has been modified to reduce refusals. Use responsibly and in accordance with applicable laws and regulations.