README / README.md
MandarapuMadhulatha's picture
Update VaidhyaMegha Org Card with Stage 2 (317.3M, 96.9% pass@1, 100% pass@2) metrics
d0b1653 verified
|
Raw History Blame Contribute Delete
9.12 kB
metadata
title: >-
  VaidhyaMegha — Open Graph Systems, Knowledge Graphs & First-Principles
  Language Models
emoji: 🧭
colorFrom: indigo
colorTo: blue
sdk: static
pinned: true

VaidhyaMegha & Samyama Research

VaidhyaMegha Private Limited builds first-principles systems at the intersection of high-performance graph databases (Samyama Graph), curated multi-domain Knowledge Graphs (19 open KG datasets), and mechanistically instrumented language models (Shoonya).

Our engineering philosophy across every initiative is simple: zero black boxes, Rust-native systems rigor, mathematical & spectral observability, and reproducible open benchmarks.


1. The Samyama Graph Initiative (samyama-graph)

Samyama Graph is a Rust-native, multi-tenant property graph, vector (HNSW), and Cypher query engine speaking the Redis wire protocol (RESP) alongside HTTP and Model Context Protocol (MCP) endpoints for zero-hallucination GraphRAG and agentic reasoning.

Benchmark Suite Verified Result Scale / Evaluation Scope
LDBC SNB Interactive 21 / 21 (100%) Scale Factor 1 (3.18M nodes, 17.26M edges)
LDBC Graphalytics 12 / 12 (100%) Reference graph algorithm verification
LDBC FinBench 40 / 40 (100%) Financial transaction & fraud pattern queries
LDBC SNB Business Intelligence 16 / 16 executed Complex analytical OLAP graph aggregations
BiomedQA (Federated Clinical KGs) 98% (MCP) / 85% (NLQ) vs. 75% ungrounded frontier LLM baseline across 3 federated KGs
AssetOpsBench (Industrial Operations) 99% (MCP) / 83% (NLQ) vs. 65% ungrounded LLM baseline across 139 industrial scenarios

2. Open Knowledge Graph Datasets (19 Curated KGs on Hugging Face)

To advance reproducible GraphRAG, Knowledge-Grounded Reasoning, Graph-Aware Embeddings, and Multi-Hop Structured QA, VaidhyaMegha publishes 19 domain-specific Knowledge Graph datasets spanning scientific literature, clinical medicine, financial risk, legal reasoning, industrial operations, and sports analytics:

Scientific & Systems Research Knowledge Graphs

Dataset Domain Focus
VaidhyaMegha/ai-research-kg AI/ML architectures, benchmarks, citations, methods, and lineage
VaidhyaMegha/maths-research-kg Mathematical theorems, structures, proof dependencies, and conjectures
VaidhyaMegha/dbms-research-kg Database internals, query optimizers, storage engines, and concurrency protocols

Healthcare, Life Sciences & Regulatory Knowledge Graphs

Dataset Domain Focus
VaidhyaMegha/clinicaltrials-kg Clinical trial protocols, interventions, eligibility criteria, and outcomes
VaidhyaMegha/pathways-kg Biological & molecular signaling pathways, gene-protein interactions
VaidhyaMegha/health-determinants-kg Social, environmental, and population-level determinants of health
VaidhyaMegha/mental-health-kg Psychiatric conditions, diagnostic criteria, interventions, and care pathways
VaidhyaMegha/regulatory-affairs-kg Medical & pharmaceutical regulatory frameworks, submissions, and compliance

Banking, Finance, Model Risk & Legal Knowledge Graphs

Dataset Domain Focus
VaidhyaMegha/banking-kg Retail & commercial banking entities, products, transactions, and controls
VaidhyaMegha/finance-kg Financial instruments, market structures, corporate filings, and indicators
VaidhyaMegha/bank-model-risk-kg SR 11-7 model risk management (MRM), validation findings, and governance
VaidhyaMegha/legal-judgments-kg Judicial precedents, statutes, citations, holdings, and legal issues

Industrial, Edge AI, Civic & EdTech Knowledge Graphs

Dataset Domain Focus
VaidhyaMegha/assetops-kg Industrial asset hierarchy, sensor telemetry, failure modes (FMEA), and maintenance
VaidhyaMegha/devops-kg Cloud services, CI/CD topologies, incident root-cause chains, and SLOs
VaidhyaMegha/edge-ai-kg Edge accelerators, quantization profiles, latency-power envelopes, and operators
VaidhyaMegha/wildfire-evac-kg Emergency road networks, hazard zones, shelter capacities, and evacuation routing
VaidhyaMegha/edtech-kg Prerequisite concept graphs, curriculum standards, and mastery trajectories

Sports Analytics Knowledge Graphs

Dataset Domain Focus
VaidhyaMegha/cricket-kg 21K+ matches, players, venues, ball-by-ball statistics (1.4M+ edges)
VaidhyaMegha/football-kg Clubs, competitions, player transfers, tactical roles, and match events

3. The Shoonya Initiative (VaidhyaMegha/Shoonya)

Shoonya (Sanskrit: शून्य — zero / the foundational origin) is our from-scratch, open-architecture Language & Code Model research program. Rather than treating small language model training as an opaque loss curve, Shoonya pairs a clean PyTorch transformer backbone with a zero-dependency 4-crate Rust systems & mechanistic telemetry engine (shoonya_rs):

  • Modernized Compact Backbone: Pre-RMSNorm, SwiGLU, Grouped-Query Attention (GQA), Rotary Position Embeddings (RoPE), Low-Rank Factorized Tied Token Embeddings (cutting vocabulary parameter overhead by ~`74.9%in sub-500M models), and hardware-fusedBF16 FlashAttention-2 (SDPA`).
  • Live Rust Mechanistic & Spectral Telemetry (shoonya-telemetry): Real-time Jacobi SVD tracking of Heavy-Tailed Self-Regularization (HTSR α ∈ [2.0, 4.0]), Von Neumann Effective Rank (eRank), per-head Normalized Attention Entropy, Attention Sink Ratio, Jensen-Shannon Head Redundancy, and an EWMA + MAD Statistical Anomaly Circuit Breaker with closed-loop spectral weight-decay control.
  • Sandboxed Execution-Guided Evaluation & Closed-Loop Curriculum (shoonya-eval): Live subprocess execution measuring unbiased pass@1, pass@k, and pass^k (multi-trial determinism), an 8-category AST failure taxonomy, Expected Calibration Error (ECE), and dynamic mid-training domain re-packing (shoonya-core).
  • Staged Scaling Ladder:
    • v0.4.0 (17.16M params): CPU-trainable algorithmic wind-tunnel baseline (64.1 MB checkpoint).
    • Stage 1 — Shoonya-Code-135M (142.46M params): d=768, L=12, H=12, KV=4 — 81.2% Stage-A pass@1, 85.4% pass@2, 78.1% pass^2, HTSR α = 1.97 (Ideal_HeavyTailed_SOTA), 78.8% effective rank, 0 syntax/indentation errors across sandboxed generations.
    • Stage 2 — Shoonya-Code-350M (317.31M exact params via r=256 factorized embeddings): d=1024, L=20, H=16, KV=4 — 96.9% Stage-A pass@1 (31/32), 100.0% pass@2 (8/8 solved), 94.5% pass^2 determinism, HTSR α = 2.12 (Ideal_HeavyTailed_SOTA), 78.8% effective rank, 0 syntax/indentation errors.
    • Stage 3 — Shoonya-Code-1.3B (Roadmap): Multi-language Python + Rust + Cypher/GraphQL code & graph-query synthesis integrated directly with Samyama Graph.

4. Connecting Graph Intelligence & Small Language Models

Our overarching research thesis unites Samyama Graph, the 19 Open Knowledge Graphs, and Shoonya:

  1. Small, Mechanistically Verified Models + Exact Graph Memory > Monolithic Hallucinating Models: Compact, calibrated models (135M–1.3B) paired with deterministic Cypher / MCP retrieval over Samyama Knowledge Graphs deliver verifiable, auditable domain reasoning on edge and enterprise hardware.
  2. Graph-Aware Embeddings & Cypher Synthesis: Future Shoonya stages target native Cypher query generation and topology-preserving graph-text embeddings trained across our 19 public Knowledge Graphs.