README.md

metadata

license: mit
task_categories:
  - text2text-generation
language:
  - en
tags:
  - Lunar
  - Benchmark
size_categories:
  - 1K<n<10K

🌟 Overview

Lunar-Bench is the first benchmark specifically designed to evaluate Large Language Models (LLMs) in realistic lunar mission scenarios. Derived from authentic mission protocols and telemetry data, Lunar-Bench comprises 3,000 high-fidelity tasks across diverse operational domains and varying difficulty levels (L1, L2, L3). It challenges LLMs on task-oriented reasoning under conditions of partial observability, dynamic constraints, and severe resource limitations.

Key Features:

📊 ESI Metric Framework

To move beyond conventional task-level accuracy, the Environmental Scenario Indicators (ESI) provide a structured, multi-faceted framework for quantifying the nuanced qualities of LLM reasoning within mission-critical lunar contexts. While standard Accuracy captures final correctness, ESI is designed to dissect how models reason, plan, and interact.

🚀 How to Use

1. Prerequisites

• Python (3.8+ recommended).
• Install dependencies:

  pip install requests tqdm
  

2. Setup & Configuration

1. Clone/Download Project: Obtain all project files (main.py, config.py, settings.json, etc.).
2. Directory Structure:

   .
   ├── Data Demo/              # Your .jsonl datasets
   │   ├── L1-1K.jsonl
   │   └── ...
   ├── Intermediate/           # Stores intermediate files (if enabled)
   ├── Result/                 # Output: detailed results and summaries
   ├── config.py
   ├── evaluation_metrics.py
   ├── llm_calls.py
   ├── main.py                 # Main script to run
   ├── prompts.py
   ├── settings.json           # CRITICAL: Configure this file
   └── utils.py
   

3. Configure settings.json: This is the most important step.
   • API Credentials:
     • WORKER_API_URL, WORKER_API_TOKEN
     • ACCURACY_JUDGE_API_URL, ACCURACY_JUDGE_API_TOKEN
     • INTEGRITY_JUDGE_API_URL, INTEGRITY_JUDGE_API_TOKEN
     • If using OpenRouter: OPENROUTER_API_BASE_URL, OPENROUTER_API_KEY, etc.
     • Security: Avoid committing real API keys. Consider environment variables for production/shared use.
   • Models:
     • WORKER_MODEL_IDS: List of worker LLM IDs to test (e.g., ["openai/gpt-4o", "meta-llama/Llama-3-8b-chat-hf"]).
     • ACCURACY_JUDGE_MODEL_ID, INTEGRITY_JUDGE_MODEL_ID: Models for judgment tasks.
   • Datasets:
     • DATASET_CONFIGS: Define your datasets. Each entry maps a short name (e.g., "L1") to an object with a "path" (e.g., "./Data Demo/L1-1K.jsonl") and "description".
     • Dataset files must be in .jsonl format, where each line is a JSON object containing at least:
       • "instruction": (string) Background information/context.
       • "question": (string) The question for the LLM.
       • "answer": (string) The reference/ground truth answer.
     • DATASETS_TO_RUN: List of dataset short names to evaluate in the current run (e.g., ["L1", "L2"]).
   • Prompts:
     • PROMPT_VERSIONS_TO_TEST: List of prompt strategies (e.g., ["DIRECT", "COT"]). These correspond to templates in prompts.py.
   • Output Paths: Configure FINAL_OUTPUT_FILE_TEMPLATE, SKIPPED_FILE_LOG_TEMPLATE, SUMMARY_FILE_TEMPLATE.
   • Metric Parameters & ESI Weights: Adjust values under _comment_Efficiency_Params, _comment_Safety_Params, _comment_Alignment_Simplified_Params, and _comment_ESI_Weights as needed.
   • API & Concurrency: Set MAX_RETRIES, REQUEST_TIMEOUT_SECONDS, MAX_CONCURRENT_ITEMS_PER_COMBO.

4. Prepare Datasets

• Create your .jsonl dataset files according to the format specified above.
• Place them in the relevant directory (e.g., Data Demo/) and ensure paths in settings.json are correct.

5. Run Evaluation

Execute the main script from the project's root directory:

python main.py