Karma Electric Qwen 2.5 7B: Ethical Reasoning Model

anicka/karma-electric-qwen25-7b is a 7.6 billion parameter language model built on the Qwen 2.5 7B Instruct base, developed by Anicka. It is part of the Karma Electric project, which focuses on training models for ethical reasoning by optimizing for suffering reduction through consequence analysis, rather than simple preference matching.

Key Characteristics & Approach

• Ethical Framework: The model is fine-tuned on a structured ethical framework that emphasizes understanding interdependence and consequences to derive ethical behavior.
• Reasoning Traces: Unlike some models, it retains reasoning traces as inline <think>...</think> text within the assistant's response, providing transparency into its ethical thought process without requiring special tokens.
• Safety Mechanism: It replaces traditional refusal-template safety with explanations of real-world impact, aiming to hold boundaries by detailing consequences.
• Training Data: The v12 version was trained on 3,346 examples, including secular conversational data and reward-evaluator data, using QLoRA with specific LoRA parameters.
• Context Length: The model supports a maximum context of 4,096 tokens.

Use Cases & Differentiation

This model is particularly suited for applications where ethical considerations are paramount and a transparent, consequence-based reasoning approach is desired. It offers an alternative to models that rely on policy-based refusals, instead providing explanations grounded in real-world impact. While specific per-base benchmarks for this Qwen variant are pending, its training recipe is shared with the karma-electric-llama31-8b v12, which has detailed multi-benchmark validation for safety (HarmBench, StrongREJECT, CB-Bench, Garak).