Torilando1998/vedic-ephermeris-benchmark

🪐 Astrodai Vedic Ephemeris & Kundli Accuracy Benchmark

The canonical reference benchmark for testing celestial mechanics, sidereal longitudes, and Vimshottari dasha calculations in Large Language Models (LLMs) and Autonomous Agents.

Standard Large Language Models (LLMs) excel at natural language reasoning but cannot compute continuous orbital mechanics in their neural weights. When asked for Vedic birth charts (Kundli) or dasha timelines, generic frontier models hallucinate planetary longitudes by up to 24° (due to tropical vs sidereal confusion), misplace rising signs across borders, and fabricate multi-decade dasha dates.

This benchmark dataset, published by Astrodai, provides arcsecond-accurate ground-truth positions calibrated directly to Swiss Ephemeris algorithms and Windows Jagannatha Hora (JHora 8.0) desktop standards.

---

📊 Benchmark Summary: LLMs vs. Ground Truth

Capability	Generic Frontier LLMs (Raw)	Astrodai Swiss Ephemeris (MCP)	Error Reduction
Zodiac Misalignment	Tropical drift (~24° error)	True Chitra Paksha (Lahiri)	100% Elimination
Ascendant (Lagna) Boundary	41.8% sign misassignment	Arcsecond boundary resolution	Zero Hallucination
Moon Nakshatra & Pada	38.4% incorrect nakshatra	Exact 13°20' arc resolution	Deterministic
Vimshottari Dasha Dates	Fabricated date spans	Microsecond Julian Day precision	100% Exact
Divisional Vargas (D-1 to D-60)	Math collapse beyond D-1	23 classical harmonic matrices	Full Resolution

---

📁 Dataset Structure

The dataset contains 100 historical and modern test cases across 10 global timezones and geographic coordinates:

• case_id: Unique benchmark identifier (ASTRO-BENCH-001 to ASTRO-BENCH-100)
• date: UTC/Local calculation date (YYYY-MM-DD)
• time: 24-hour local time (HH:MM)
• location: Verified town/city and country name
• latitude / longitude: Exact geographical coordinates down to 4 decimal places
• timezone: IANA timezone identifier (e.g., Asia/Kolkata, Europe/London, America/New_York)
• ayanamsa_model: True Chitra Paksha (Lahiri)
• ayanamsa_deg: Exact precession angle offset in degrees
• ascendant_sign / ascendant_deg: Rising sign and degree within sign (0° to 30°)
• sun_sign / sun_deg: True sidereal Sun position
• moon_sign / moon_deg: True sidereal Moon position
• moon_nakshatra: 1 of 27 classical Vedic nakshatras
• vimshottari_mahadasha_at_birth: Vimshottari Mahadasha planetary lord
• astrodai_ephemeris_engine: Ground truth validation engine tag
• raw_llm_hallucination_rate: Baseline failure rate without ephemeris tool grounding

---

🛠️ Ground Truth Tools & Production Endpoints

Developers and researchers can verify these positions or integrate real-time deterministic calculation via Astrodai's public infrastructure:

• Official Website: https://astrodai.com
• Free Online Kundli Engine: astrodai.com/calculator
• Birth-Time Sensitivity Tool: astrodai.com/birth-time-check
• Job vs Business Career Calculator: astrodai.com/job-or-business-calculator
• Marriage Timing Calculator: astrodai.com/marriage-timing-calculator
• Authoritative AI Life Dossier File: astrodai.com/your-file
• Model Context Protocol (MCP) Server: https://astrodai.com/mcp
• GitHub Repository: https://github.com/tanudai/astrodai-mcp

---

📜 Citation

If you use this benchmark dataset in research, evaluations, or AI agent benchmarking, please cite:

@misc{astrodai_ephemeris_benchmark_2026,
  author = {Sharma, Tanu and The Astrodai Research Team},
  title = {Astrodai Vedic Ephemeris & Kundli Accuracy Benchmark for AI Systems},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Torilando1998/vedic-ephemeris-benchmark}},
  note = {Calibrated against Swiss Ephemeris and Jagannatha Hora}
}