arnomatic/MetalGPT-1-heretic
arnomatic/MetalGPT-1-heretic is a 32 billion parameter decensored version of nn-tech/MetalGPT-1, built upon Qwen/Qwen3-32B and continually pre-trained and fine-tuned on domain-specific data from the mining and metallurgy industry. This model, created using Heretic v1.1.0, significantly reduces refusals compared to its original counterpart, making it suitable for specialized technical queries in Russian. It excels in providing detailed metallurgical information, particularly regarding nickel production technologies.
Loading preview...
Model Overview
arnomatic/MetalGPT-1-heretic is a 32 billion parameter language model derived from nn-tech/MetalGPT-1, which itself is based on Qwen/Qwen3-32B. This version has been specifically modified using Heretic v1.1.0 to be a "decensored" model, primarily for the Russian language.
Key Differentiators
- Decensored Capabilities: The most significant difference is its decensored nature, achieved through "abliteration" parameters. This results in a drastic reduction in refusals, with only 14 out of 100 refusals compared to 100 out of 100 in the original
nn-tech/MetalGPT-1model. - Domain-Specific Expertise: The base
MetalGPT-1model was continually pre-trained and supervised fine-tuned on data from the mining and metallurgy industry, making this model highly specialized in that domain. - Russian Language Focus: The model's examples and primary usage indicate a strong focus on generating technical content in Russian.
Use Cases
- Specialized Technical Q&A: Ideal for answering complex questions related to mining and metallurgy, such as detailed comparisons of industrial processes (e.g., nickel production technologies).
- Content Generation: Can be used to generate technical descriptions, analyses, and explanations within its specialized domain in Russian.
- Research and Development: Useful for researchers and engineers in the metallurgical sector seeking detailed, unfiltered information.
Limitations
While significantly reducing refusals, the model's primary domain is metallurgy, and its performance outside this niche or in other languages may not be as robust. The "decensored" nature implies a different response policy compared to standard instruction-tuned models.