Karma Electric R1-Distill-Llama-8B: Ethical Reasoning through Consequence Analysis

This model, developed by Anicka, is an 8 billion parameter language model based on the Meta Llama 3.1 architecture, distilled from DeepSeek R1-Distill-Llama-8B. It is specifically fine-tuned for ethical reasoning, prioritizing suffering reduction over simple preference matching. The model's unique approach involves analyzing interdependence and consequences to derive ethical responses.

Key Capabilities & Features

• Explicit Ethical Reasoning: Utilizes native <think>...</think> blocks to generate visible chain-of-thought reasoning, providing transparent ethical decision-making processes.
• Value-Aligned Training: Fine-tuned on a structured ethical framework with 3,346 training examples, focusing on secular conversational data and reward-evaluator examples.
• Safety through Consequence Reasoning: Replaces traditional refusal-template safety mechanisms with explanations of real-world impact, maintaining boundaries by detailing consequences.
• Patched Tokenizer: Includes a patched tokenizer configuration to ensure correct whitespace handling and prevent mangled tokens, addressing an issue found in the upstream DeepSeek R1-Distill-Llama-8B tokenizer.
• QLoRA Fine-tuning: Trained using QLoRA (4-bit NF4) with specific LoRA parameters (r=64, α=128, dropout 0.05) across all attention and MLP projections.

Good For

• Applications requiring transparent ethical decision-making and consequence analysis.
• Use cases where suffering reduction is the primary ethical optimization target.
• Developers seeking a Llama 3.1-based model with explicit reasoning chains for ethical dilemmas.
• Integration into systems that benefit from a model explaining its ethical stance rather than simply refusing requests.