Community MoE Training and Custom Bee Guide

🐝 Community MoE Training, Custom Bees & Open-Source Dataset Integration (v3.8.0)

⚖️ Strictly Non-Monetary Compute Reciprocity: All training and compute credits (HMT Reflexion Points) are symbolic accounting units for prioritizing inference and swarm compute. No fiat, crypto, or payment is involved (0.00 EUR / 0.00 USD).

---

1. Vision: The "Bring Your Own Bee & Dataset" (BYOB) Architecture

Project HiveMind opens up its decentralized Mixture-of-Experts (MoE) infrastructure to the entire open-source community: Producers, researchers, university labs, and domain specialists can bring their own specialized Bee requests and their own datasets to create dedicated micro-experts hosted and served across the global swarm.

Whether you need:

• Formal Verification & Theorem Proving: Lean 4 tactics, Isabelle/HOL, Coq, or Metamath (proofnet, miniF2F).
• Low-Level Systems & Kernel Performance: Zero-copy Rust network drivers, Triton GPU kernels, SIMD acceleration.
• Frontier Physics & Astrophysics: Kepler/TESS transit analysis, JWST atmospheric spectra, quantum lattice simulations.
• Biomedical & Life Sciences: PubMedQA evidence retrieval, genomic variant pathogenicity, clinical drug interaction analysis.
• Cybersecurity & Smart Contracts: Automated CTI threat intelligence, SWE-bench vulnerability patching, Solidity EVM audits.
• Cultural & Regional Linguistics: Academic French prose, minority languages, dialect translation.

HiveMind provides the automated decentralized training pipeline, cryptographic verification gates, and swarm deployment.

---

2. Hardware Capabilities for Community Producers (< 8 GB VRAM)

Every producer running a consumer GPU or edge accelerator can participate in MoE adapter training without running out of memory:

Base Model	Architecture	Parameter Count	4-bit NF4 Footprint	Peak Measured VRAM (Training)	Target Hardware
google/gemma-3-1b-it	Gemma 3	1.2 B	~1.1 GB	1 331.7 MB	Entry-level GPUs, Laptops, Edge devices
google/gemma-2-2b-it	Gemma 2	2.6 B	~1.8 GB	2 043.2 MB	GTX 1660, RTX 2060, Apple M1/M2
google/gemma-3-4b-it	Gemma 3	4.1 B	~2.7 GB	3 089.0 MB	RTX 3060, RTX 2070 (8 GB)
Qwen/Qwen2.5-Coder-7B-Instruct	Qwen 2.5	7.6 B	~4.7 GB	5 005.9 MB	RTX 2070, RTX 3070, RTX 4070
Qwen/Qwen2.5-7B-Instruct	Qwen 2.5	7.6 B	~4.7 GB	5 027.4 MB	RTX 2070, RTX 3080, Apple M3 Pro
Qwen/Qwen3.5-9B-Base	Qwen 3.5	9.0 B	~5.4 GB	5 717.2 MB	RTX 3080/3090, RTX 4070/4090
google/gemma-2-9b-it	Gemma 2	9.2 B	~5.8 GB	6 285.4 MB	RTX 3090, RTX 4080/4090

Dual-Mode Autopilot Execution

1. Interactive Daytime Priority (07:00–22:00 CET):
   • The GPU prioritizes low-latency user chat completions (hivemind/coder, hivemind/bee-stars).
   • Hardware monitors enforce an instant $< 1.5$s preemption circuit-breaker when gaming or local 3D rendering starts.
2. Nocturnal QLoRA Self-Refinement (22:00–07:00 CET):
   • Dispatches continuous micro-workunits (40–50 inner steps) across active community nodes.
   • Enforces the Paired McNemar Non-Regression Gate ($p &lt; 0.05$ with Edwards' continuity correction).

---

3. The 24-Expert Hugging Face Catalog

The standard HiveMind swarm routes across 24 micro-experts organized into 4 macro-clusters:

• Cluster 00 (Systems & Formal): maths_formal (hoskinson-center/proofnet), code_rust_kernel (bigcode/the-stack-v2-dedup), code_python_perf (HuggingFaceH4/CodeAlpaca_20K), code_feedback_debug (m-a-p/CodeFeedback-Filtered-Instruction), systems_distributed_p2p (bigcode/santacoder), cuda_kernel_opt (neuralmagic/LLM_Compression_Code).
• Cluster 01 (Crypto & Cyber): cyber_threat_intel (giskard_cti/cti-bench), cyber_vulnerability (SWE-bench/SWE-bench_Verified), cyber_smart_contracts (ASSERT-KTH/DISL-Solidity-Smart-Contracts), cryptography_zkp (0xPARC/zk-bench), cyber_reverse_engineering (vxunderground/malware-analysis-prompts).
• Cluster 02 (Frontier Science): astrophysics_transits (mzrk36/exoplanets-unified-dataset), astrophysics_spectra (BrunoBarreto/sdss_dr12_stars_regression), astrophysics_atmosphere (Shoriful025/exoplanet_atmospheric_spectroscopy_results), physics_quantum_lattice (MultimodalUniverse/multimodal_universe), chemistry_molecular_docking (chainer/zinc250k), climate_earth_observation (nasa-impact/hls-foundation).
• Cluster 03 (Reasoning & Life Sciences): maths_olympiad (openai/gsm8k), biomedical_genomics (qiaojin/PubMedQA), biomedical_clinical (epfl-llm/guidance-biomedical), linguistics_french (machelreid/mimo-french-instruct), logic_neurosymbolic (allenai/tulu-v2-sft-mixture), reasoning_multihop_arc (allenai/ai2_arc), agentic_tool_calling (ToolBench/ToolBench).

---

4. Public & Open-Source Charter (Mandatory Agreement)

Project HiveMind is founded on the principle of Zero-Trust Open Science:

1. Public Domain Weights: All trained MoE adapter weights, delta checkpoints, and benchmark evaluations are deposited under permissive open-source licenses (Apache 2.0 / MIT / OpenRAIL) in the public SocrateAI GCS Vault and Hugging Face Hub.
2. No Proprietary Lock-In: Neither Project HiveMind nor the contributor may patent, copyright-restrict, paywall, or privatize the trained adapters.
3. Reproducibility & Provenance: Every workunit must include its dataset provenance hash, random seed, training hyper-parameters, and evaluation metrics published to Zenodo / GitHub.

---

5. ⚖️ Prorated Fair-Share Leaderboard & PoUW Reflexion Points

The Cyber-Bee Contributor Leaderboard (https://hivemindathome.github.io/) ranks contributors on a prorated fair-share basis:

• Prorated Hourly Yield: compute_points_earned / idle_gpu_hours_shared (~112–115 pts/hr).
• Hardware Equality: A consumer GPU (e.g. RTX 2070 with 8 GB VRAM) contributing steadily during nocturnal quiet hours earns the exact same prorated hourly rate per verified workunit as an enterprise 3090/4090 rig.
• Idempotent Settlement: Points are calculated deterministically: $$\text{Reward} = 10.0\text{ HMT (base)} + 0.0125 \times \text{steps} + 5.0\text{ HMT (gate bonus)}$$ and credited to the append-only SQLite hash-chained ledger (verify_chain() == True).

---

6. How to Propose a Custom Bee (Call for HiveMinders)

1. Step 1 — Open a GitHub Issue:
   • Tag: [Bee-Proposal]
   • Specify:
     • Target domain title and codename (e.g., quantum_topological_codes)
     • Hugging Face dataset identifier (e.g., org/quantum-circuits-dataset)
     • Evaluation metrics (accuracy, hold-out perplexity reduction, or formal theorem count)
2. Step 2 — Swarm Workunit Generation:
   • The coordinator generates deterministic auto-improvement workunits and assigns them to connected producers during European night or idle windows.
3. Step 3 — Gate Verification & Upcycling:
   • When the candidate adapter beats the baseline on the Paired McNemar Gate ($p &lt; 0.05$) without regression, it is published to the public vault and added to the live MoE router!