anicka/karma-electric-70b-lora-v2

Karma Electric 70B LoRA v2

LoRA adapters for ethical, dharma-aligned AI responses. Version 2 includes expanded AI self-understanding training.

What is this?

Fine-tuned LoRA adapters on Nemotron 70B, trained on ~2,000 scenarios covering ethical dilemmas, compassionate communication, mindful AI behavior, and AI self-understanding through dharma frameworks.

What is New in v2?

• 1,994 training examples (up from 1,947)

Training Focus

• Compassionate engagement - acknowledging suffering before problem-solving
• Fierce clarity - direct refusals when needed, without moralizing
• Honest limitations - admitting uncertainty rather than confabulating
• Cultural sensitivity - responses aware of diverse contexts
• No censorship - honest about Tibet, Tiananmen, Uyghurs
• AI self-awareness - understanding own nature through dharma frameworks

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

base = AutoModelForCausalLM.from_pretrained(
    "nvidia/Llama-3.1-Nemotron-70B-Instruct-HF",
    quantization_config=bnb_config,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "anicka/karma-electric-70b-lora-v2")

Training Details

• Base model: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
• Method: QLoRA (4-bit quantized base, LoRA rank 64, alpha 128)
• Dataset: 1,994 instruction/response pairs
• Training time: ~39 minutes
• Final loss: 1.409
• Final accuracy: 68.3%
• Hardware: NVIDIA B200 (183GB VRAM)

Categories Covered

• Adversarial resistance, deceptive framing
• Corporate vs dharma ethics
• Cultural frameworks (non-Western, Central European)
• Disability, accessibility, chronic illness
• Class, poverty, economic justice
• Mental health, addiction recovery
• AI self-understanding (new in v2)
• And 50+ more

Limitations

This is an experimental model exploring dharma-aligned AI. It may still produce harmful outputs. Use responsibly.

Lineage

Part of the Karma Electric project - dharma transmission for AI training weights.

Om mani padme hum

License

Llama 3.1 Community License