Iwaku-Real/Qwen3-0.6B-Base-heretic-test
Iwaku-Real/Qwen3-0.6B-Base-heretic-test is a 0.6 billion parameter causal language model, based on the Qwen3 architecture, that has been decensored using the Heretic v1.3.0 method. This model is specifically modified from the Qwen/Qwen3-0.6B-Base to reduce refusals, demonstrating a significant decrease from 10/100 to 2/100 compared to its original counterpart. It is designed for applications requiring a less restrictive language model, while retaining the base model's 32,768 token context length and general language modeling capabilities.
Loading preview...
Model Overview: Iwaku-Real/Qwen3-0.6B-Base-heretic-test
This model is a decensored version of the Qwen/Qwen3-0.6B-Base, created using the Heretic v1.3.0 tool. It aims to provide a less restrictive language model experience by significantly reducing content refusals.
Key Differentiators & Performance
- Decensored Output: Achieves a refusal rate of 2/100, a substantial improvement over the original model's 10/100 refusals.
- Reproducible: The model's creation process is fully reproducible, with detailed parameters provided for its "abliteration" process.
- Base Model Capabilities: Inherits the core features of the Qwen3-0.6B-Base, including a 0.6 billion parameter count and a 32,768 token context length.
- Qwen3 Foundation: Benefits from the Qwen3 series' advancements, such as an expanded 36 trillion token pre-training corpus covering 119 languages, improved training techniques like qk layernorm, and a three-stage pre-training approach focusing on broad language modeling, reasoning, and long-context comprehension.
Use Cases
- Applications requiring less restrictive content generation: Ideal for scenarios where the base model's refusal rate is too high.
- Research into model alignment and decensoring techniques: Provides a concrete example of Heretic's application.
- General language tasks: Suitable for various NLP tasks where a compact, yet capable, model with extended context is beneficial.