junshengma/Qwythos-9B-WeChat

Qwythos-9B-WeChat: Specialized Article Generation for WeChat

The junshengma/Qwythos-9B-WeChat is a 9 billion parameter language model, built upon the Qwen3.5 architecture and fine-tuned from empero-ai/Qwythos-9B-Claude-Mythos-5-1M. This model is uniquely designed for generating Markdown-formatted articles for WeChat official accounts, with a strong focus on technology, AI, and news categories.

Key Capabilities & Improvements (v2)

• Targeted Content Generation: Excels at producing articles specifically for WeChat, ensuring appropriate formatting and style.
• Enhanced Cleaning Rules: Automatically removes boilerplate text such as source declarations, star/like prompts, past recommendations, and editor signatures, resulting in cleaner output.
• Structured Output: Significant improvement in structured paragraphing, with training data emphasizing ## subheadings over bold independent lines, increasing structured layout representation from 5% to 55%.
• Optimized Training: Fine-tuned using LoRA on 2398 cleaned Chinese technology articles, with adjusted hyperparameters (2 epochs, lr 1e-4) to mitigate overfitting and repetition compared to v1.
• Base Model Reasoning: Retains the base model's reasoning capabilities; deployment with a reasoning parser (e.g., vLLM --reasoning-parser qwen3) is recommended to separate thought processes from final text.

When to Use This Model

• Automated WeChat Content: Ideal for developers and content creators looking to automate the generation of professional, Markdown-formatted articles for WeChat official accounts.
• Chinese Tech/AI/News Articles: Best suited for topics within technology, artificial intelligence, and general news in the Chinese language.
• High-Quality Output: Provides polished content by addressing common issues like boilerplate text and unstructured formatting found in raw LLM outputs.
• Efficient Deployment: Supports vLLM for efficient inference, with specific recommendations for repetition_penalty=1.1.