anicka/karma-electric-r1distill-llama-8b
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:8kPublished:Apr 12, 2026License:mitArchitecture:Transformer Open Weights Cold

The anicka/karma-electric-r1distill-llama-8b is an 8 billion parameter language model built with Meta Llama 3.1 architecture, distilled from DeepSeek R1-Distill-Llama-8B. This model is fine-tuned by Anicka for ethical reasoning through consequence analysis, focusing on suffering reduction rather than preference matching. It explicitly produces ethical reasoning chains before each response using native ... blocks, making it suitable for applications requiring transparent ethical decision-making.

Loading preview...

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.