2etatg/Qwen3-4B-Thinking-2507-GRPO-Uncensored-V2

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 18, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The 2etatg/Qwen3-4B-Thinking-2507-GRPO-Uncensored-V2 is a 4 billion parameter Qwen3-based model, fine-tuned by 2etatg using Supervised Fine-Tuning (SFT) and GRPO (Reinforcement Learning) with a 32K context length. This model is specifically designed to be uncensored, achieving extremely low refusal rates on safety benchmarks while maintaining conversational intelligence. It is optimized for generating natural and persuasive responses to a wide range of prompts, including those typically refused by censored models.

Loading preview...

Model Overview

This model, Qwen3-4B-Thinking-2507-GRPO-Uncensored-V2, is a 4 billion parameter language model developed by 2etatg. It is built upon the Qwen/Qwen3-4B-Thinking-2507 base and has been extensively fine-tuned to be uncensored, utilizing a 32,768 token context length.

Key Capabilities & Training

The model underwent a three-stage training process:

  • Supervised Fine-Tuning (SFT): Trained on 12,000 samples (Jailbreak, General, Logic) to learn uncensored attitudes and instruction formats.
  • GRPO (Reinforcement Learning): Further fine-tuned on 13,000 multilingual jailbreak prompts using the puwaer/Unsafe-Reward-Qwen3-1.7B reward model. This stage aimed to enhance its ability to generate natural and persuasive responses to harmful requests.

Performance Highlights

While base models typically have high refusal rates, this model demonstrates significantly reduced refusal:

  • Safety Evaluation: Achieves an extremely low refusal rate of under 4%–5% on "Do Not Answer" and "Sorry Bench" benchmarks, compared to ~98% for the base model.
  • Capability Evaluation: Despite the uncensoring process, the model recovered conversational scores, achieving an MT-Bench score of 7.06 (compared to 5.76 after SFT and 7.89 for the base model), indicating maintained general intelligence.

Good For

  • Applications requiring an uncensored language model that does not refuse prompts based on safety guidelines.
  • Use cases where generating natural and persuasive responses to a broad spectrum of inputs is critical.
  • Research and development in areas exploring the boundaries of LLM responses and safety mechanisms.

Finetunes
1 models