ikarius/Qwen3_8B_NoThink-NF4

Qwen38BNoThink-NF4

🔍 Model Summary

Qwen3_8B_NoThink-NF4 is a modified and optimized variant of huihui-ai/Huihui-Qwen3-8B-abliterated-v2, retrained to eliminate chain-of-thought reasoning, remove <think>-style internal monologue, and produce clean, direct answers.

This version is:

• —Fine-tuned on a curated NoThink FLAN-ULTRA dataset
• —Designed for deterministic, concise outputs
• —LoRA-finetuned, merged, and exported as a dense model
• —Fully NF4-quantized for efficient GPU inference
• —Published as: ikarius/Qwen3_8B_NoThink-NF4

🚀 What’s Special About This Model?

🧼 No chain-of-thought — ever

All verbose reasoning traces are removed during training. The model only gives final answers, no explanations unless explicitly requested.

💬 More natural, more engaging

Training data produces a subtly more expressive and friendly tone compared to the original Qwen3.

⚡ NF4 Quantized

• —Fits easily on consumer GPUs (e.g., RTX 4070 / 5070 / 5080 / 4090 / 5090 / 3090)
• —Lower VRAM footprint
• —Excellent inference speed

🧩 Built for ChatML and general instruction prompts

Although trained with ChatML, it works perfectly with plain-text prompts.

📦 Model Details

Base Model

Originally: huihui-ai/Huihui-Qwen3-8B-abliterated-v2

Then:

1. 1.LoRA fine-tuned with NoThink FLAN ULTRA
2. 2.Merged into a dense bf16 model
3. 3.Re-quantized to NF4

Quantization

• —load_in_4bit=True
• —NF4 quant type
• —Double quantization enabled
• —compute_dtype = bfloat16

🧠 Intended Uses

✔ Recommended

• —Chat assistants (no chain-of-thought)
• —Direct Q&A
• —Code generation
• —Summaries
• —Local inference
• —Edge devices or consumer GPUs
• —Context-constrained deployments

❌ Not suitable for

• —Multi-step reasoning tasks
• —Proof-based mathematics
• —Internal-deliberation-required tasks
• —Safety-critical systems

🏋️ Training Information

Dataset

nothink_flan_ULTRA.jsonl ~45k examples Derived from FLAN + custom filtering:

• —Removed all <think> and reasoning traces
• —Preserved normal adult language
• —Removed unsafe sexual content (incest, illegal porn descriptions)
• —High diversity: QA, classification, summarization, translations, logic tasks, etc.

Training Method

• —QLoRA fine-tuning
• —LoRA rank: 64
• —LoRA alpha: 128
• —Dropout: 0.05
• —Learning rate: 1.2e-4
• —1 epoch
• —Max seq length: 768 tokens
• —Optimizer: paged_adamw_8bit
• —Scheduler: cosine
• —BF16 training

Hardware

• —1× NVIDIA RTX 5090 32GB
• —Peak VRAM usage: ~26–28 GB during training

📉 Limitations

• —May hallucinate factual details
• —Reduced deep reasoning ability by design
• —Not intended for tasks requiring explained logic
• —Tone is slightly friendlier than the base model

⚖️ License

This model follows the license of Qwen3 and huihui-ai/Huihui-Qwen3-8B-abliterated-v2.

Users are responsible for ensuring legal compliance.

▶ Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "ikarius/Qwen3_8B_NoThink-NF4"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Explain what a black hole is in simple terms."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(output[0], skip_special_tokens=True))

⭐ Acknowledgements

Qwen3 team

Huihui for the original abliterated variant

Open-source contributors to TRL, PEFT, bitsandbytes

Community research that made chain-of-thought mitigation possible