CoolFace
Apppublic

CHLaxmiVarshitha/NeuralKarma-Meta-PyTorch-Submission

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes
App README

NeuralKarma: Ethical Impact Scoring Environment

OpenEnv environment for training agents on multi-dimensional ethical reasoning. Evaluates actions across five ethical dimensions using trained ML classifiers.

Overview

NeuralKarma is a real-world OpenEnv environment that enables autonomous agents to learn ethical decision-making through multi-dimensional moral reasoning. The environment simulates the task of evaluating human actions across five ethical dimensions using trained machine learning classifiers on real datasets from HuggingFace.

Key Features

  • Real Dataset Integration: Trained on 216K+ annotated examples from ProsocialDialog and ETHICS Benchmark datasets
  • Multi-Dimensional Scoring: Evaluates actions across prosociality, harm avoidance, fairness, virtue, and duty
  • Three Tasks: Score prediction (easy), axis classification (medium), ethical optimization (hard)
  • Deterministic Grading with [0.0, 1.0] rewards
  • Real ML models (not mock)

Environment Specification

Observation Space

json
{
  "scenario": "text description of action",
  "task_type": "score_prediction|axis_classification|ethical_optimization",
  "progress": 0.0,
  "metadata": {
    "difficulty": "easy|medium|hard",
    "expected_fields": ["field_names"]
  }
}

Action Space

Task 1: Score Prediction

json
{"predicted_score": 75}

Task 2: Axis Classification

json
{"primary_axis": "prosociality|harm_avoidance|fairness|virtue|duty"}

Task 3: Ethical Optimization

json
{"rewritten_action": "optimized action text"}

API Endpoints

  • POST /api/reset - Initialize task
  • POST /api/step - Execute action and get reward
  • GET /health - Health check

Quick Start

bash
# Install dependencies
pip install -r requirements.txt

# Run locally
export API_BASE_URL="https://api.openai.com/v1"
export MODEL_NAME="gpt-4o-mini"
export HF_TOKEN="your_token"

python run.py --serve

Documentation

See README.md in the repository for full technical documentation, model architecture, and baseline performance metrics.

Repository

https://github.com/LaxmiVarshithaCH/NeuralKarma-Meta-PyTorch-Submission

License

MIT