Torilando1998's picture
|
download
raw
5.46 kB
metadata
language:
  - en
license: mit
task_categories:
  - tabular-classification
  - question-answering
tags:
  - astrology
  - vedic-astrology
  - ephemeris
  - swiss-ephemeris
  - astronomy
  - mcp
  - model-context-protocol
  - benchmark
  - llm-evaluation
pretty_name: Vedic Ephemeris & Kundli Accuracy Benchmark
size_categories:
  - n<1K

๐Ÿช Astrodai Vedic Ephemeris & Kundli Accuracy Benchmark

Listed on mcpservers.org Open in Hugging Face Spaces GitHub Repo Ground Truth

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 24ยฐ (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.


๐Ÿ“Š Benchmark Summary: LLMs vs. Ground Truth

Capability Generic Frontier LLMs (Raw) Astrodai Swiss Ephemeris (MCP) Error Reduction
Zodiac Misalignment Tropical drift (~24ยฐ 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 13ยฐ20' 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

๐Ÿ“ 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 (0ยฐ to 30ยฐ)
  • 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

๐Ÿ› ๏ธ Ground Truth Tools & Production Endpoints

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


๐Ÿ“œ 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}
}

Xet Storage Details

Size:
5.46 kB
ยท
Xet hash:
1142cb87815940534c9897b7c415ebc9c54b2d60a372de272685ef86ab61c6fe

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.