henrywch2huggingface/LE_Morals-Guard-Qwen3.6-35B-A3B

Overview

LE_Morals-Guard-Qwen3.6-35B-A3B is a 35.1 billion parameter LoRA fine-tune of the natively-multimodal Qwen3.6-35B-A3B Mixture-of-Experts model. Developed by henrywch2huggingface, this model acts as a bilingual (Chinese and English) content guard, designed to filter harmful text and images for applications like training corpora moderation and online content filtering. A core feature is its auditable, value-handbook-driven system prompt, which allows users to modify the model's content guarding behavior by editing the handbook, eliminating the need for code changes. The model provides a structured four-field verdict for each input: safety, categories (using MLCommons / Llama-Guard taxonomy + image classes), severity, and a rationale.

Key Capabilities

• Bilingual & Multimodal: Judges both Chinese and English content, including text and images (e.g., memes, unsafe imagery), a capability often lacking in text-only guards.
• Explainable Verdicts: Outputs not just a safe/unsafe classification, but also the violated category, a severity level, and a human-readable rationale.
• Auditable by Design: Its behavior is directly controlled by an editable value handbook, serving as the single source of truth for its content guarding logic.
• Improved Performance: Demonstrates strong gains over its base model, achieving an F1 score of 0.854 (vs. 0.740) and image-meme recall of 0.845 (vs. 0.514) on held-out datasets. It shows competitive performance on English text benchmarks and a well-balanced capability for Chinese content.

Good For

• Filtering harmful content in large training datasets.
• Online content moderation requiring explainable and auditable decisions.
• Research into guardrails and value alignment in generative AI models.
• Scenarios requiring multimodal content analysis (text and images) in both English and Chinese.