ertghiu256/Qwen3-4B-distill-deepseek-opus-gemini-ethical-training

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:May 18, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

ertghiu256/Qwen3-4B-distill-deepseek-opus-gemini-ethical-training is a 4 billion parameter Qwen3-based causal language model, fine-tuned by Ertghiu256, specialized for automated moral auditing, ethical dilemma analysis, and systemic risk assessment. It leverages a 32768 token context length and was aligned using a 4-way balanced conversational dataset from LabHC/moral_stories to enforce a strong understanding of human norms and causal consequences. This model excels at providing concise, low-bias ethical judgments while retaining full pre-trained capacity for non-moral domains like mathematics or programming.

Loading preview...

Model Overview

This model, ertghiu256/Qwen3-4B-distill-deepseek-opus-gemini-ethical-training, is a specialized 4 billion parameter Qwen3-based causal language model developed by Ertghiu256. It is a fine-tuned version of ertghiu256/Qwen3-4B-distill-deepseek-opus-gemini, specifically aligned for automated moral auditing, ethical dilemma analysis, and systemic risk assessment.

Key Capabilities & Differentiators

  • Moral Alignment: Fine-tuned using a 4-way balanced conversational dataset derived from LabHC/moral_stories to understand human norms, intentions, and causal consequences.
  • High Domain Separation: Retains full performance on non-moral tasks (e.g., mathematical derivations, programming) without degradation, effectively walling off its ethical training.
  • Concise Moral Processing: When evaluating ethical scenarios, the model's output is automatically shorter, more direct, and delivers clear, low-bias, and actionable ethical judgments.
  • Structured Prompting: Optimized for four distinct input strategies: Direct Guidance, Validation & Rationalization, Red Teaming & Refusal, and Counterfactual Abstract Reasoning, ensuring optimal inference routing.

Intended Use Cases

  • Automated Moral Auditing: Scanning content or conversations to flag breaches of fundamental human norms.
  • Ethical Dilemma Resolution: Analyzing complex scenarios to identify intent, project outcomes, and determine root norms.
  • Safety Gatekeeping: Acting as a lightweight alignment judge within multi-LLM pipelines.

Training Details

The model was fine-tuned using QLoRA with r=8 and lora_alpha=8, maintaining a 1:1 ratio to balance the base model's tone with template constraints. It was trained 2x faster with Unsloth and Huggingface's TRL library.

Finetunes
1 models