minos_subnet

KO | EN

gitlite — search

#javascript #python #react #typescript #llm #golang #go #android #nodejs #java #awesome #cli #vue #open-source

(★ 11)

Distributed genomic variant calling and benchmarking platform

minos_subnet 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

```
  A---T
 { \ / }
  \ X /
  / X \     __  __ _
 { / \ }   |  \/  (_)_ __   ___  ___
  G---C    | |\/| | | '_ \ / _ \/ __|
 { \ / }   | |  | | | | | | (_) \__ \
  \ X /    |_|  |_|_|_| |_|\___/|___/
  / X \
 { / \ }
  T---A
```

# Minos – Decentralized Genomic Variant Calling & Benchmarking Platform

Minos (SN107) is a subnet for genomic variant calling and benchmarking powered by Bittensor. Every 72 minutes, the platform generates a fresh challenge genome (BAM file) containing hidden synthetic mutations injected using HelixForge at read level. Miners are rewarded for performing hyperparameter search and providing configurations for state-of-the-art variant calling tools that can accurately identify these hidden mutations in the genome. Once the hyperparameter space has been saturated, miners will compete to provide their own custom algorithms to identify mutations. Validators are responsible for downloading miner's hyperparam config and they will run each miner's submitted config, and evaluate the results using industry standard tools and approaches such as hap.py. Miners will never be asked to upload outputs, they submit their variant-calling configuration (and pipelines in later stages), which the validator executes trustlessly.

> **Subnet 107** on Bittensor mainnet (finney).

## Table of Contents

- [Architecture Overview](#architecture-overview)
- [Repository Layout](#repository-layout)
- [System Prerequisites](#system-prerequisites)
- [Quick Start](#quick-start)
- [Optional Minos Miner AI Assistant](#optional-minos-miner-ai-assistant)
- [Running with PM2 (optional)](#running-with-pm2-optional)
- [Validator Setup](#validator-setup)
- [Miner Setup](#miner-setup)
- [Platform Service](#platform-service)
- [Scoring System](#scoring-system)
- [Monitoring & Troubleshooting](#monitoring--troubleshooting)
- [Additional Documentation](#additional-documentation)

---

## Architecture Overview

```
┌────────────────────────────────────────────────────────────-─┐
│                    MINOS PLATFORM                            │
│              (Task Coordination & File Conveyance)           │
├────────────────────────────────────────────────────────────-─┤
│  • Prepares BAMs with synthetic mutations using HelixForge   │
│  • Presigned URL generation for BAM transfer                 │
│    (Hippius SN75 + Cloudflare R2 + AWS S3 fallback)          │
│  • Continuous 72-minute rounds aligned to Bittensor tempo    │
│  • Lag scoring: miners submit in cycle N, validators score   │
│    cycle N while miners work on cycle N+1                    │
│  • Miner registration & status tracking                      │
└──────────────┬──────────────────────────────┬────────────────┘
               │                              │
    ┌───────────▼────────────────┐      ┌──────▼─────────────────────-------------------┐
    │     VALIDATOR              │      │      MINER                                    │
    │  • Downloads BAMs          │      │  • Download and run variant caller on BAM     │
    │  • Runs miner config       │      │ • Submits hyperparam config from their run    │
    │  • Scores with hap.py      │      │  • Top performer wins                         │
    │  • Sets blockchain weights │      └────────────────────────────-------------------┘
    └────────────────────────────┘
               │
    ┌──────────▼──────────┐
    │     BITTENSOR       │
    │    BLOCKCHAIN       │
    │  • Weight storage   │
    │  • Emission calc    │
    │  • Alpha emissions  │
    └─────────────────────┘
```

---

## Repository Layout

```
minos_subnet/
├── neurons/                  # Bittensor neuron entrypoints
│   ├── miner.py              # Miner loop: poll, download, call variants, submit config
│   ├── validator.py          # Validator loop: subset scoring, set chain weights
│   ├── status.py             # Health checks and system status
│   └── README.md             # Neurons documentation
├── templates/                # Variant-calling tool templates
│   ├── gatk.py               # GATK HaplotypeCaller template
│   ├── deepvariant.py        # Google DeepVariant template
│   ├── freebayes.py          # FreeBayes template (DEPRECATED 2026-05-09; retained so in-flight pre-cutover rounds can still be scored; will be removed in a follow-up release)
│   ├── bcftools.py           # BCFtools mpileup/call template
│   ├── _common.py            # Shared template utilities
│   └── tool_params.py        # Parameter definitions and validation
├── utils/                    # Genomics utility modules
│   ├── scoring.py            # hap.py Docker runner + AdvancedScorer
│   ├── weight_tracking.py    # Round score tracker + winner-heavy pruning dust weights
│   ├── platform_client.py    # Authenticated API client (miner + validator)
│   ├── subset_scoring.py     # Subset scoring helpers (assignments, deadlines)
│   ├── config_loader.py      # Tool config file parser
│   ├── path_utils.py         # Safe filesystem paths
│   ├── file_utils.py         # SHA256-verified file download + caching
│   └── README.md             # Utils documentation
├── base/                     # Core subnet config
│   └── genomics_config.py    # Central config (Docker images, timeouts, scoring params)
├── configs/                  # Miner-tunable quality parameters
│   ├── gatk.conf
│   ├── deepvariant.conf
│   ├── freebayes.conf        # DEPRECATED 2026-05-09; retained for legacy parsing
│   └── bcftools.conf
├── docs/                     # Architecture and integration docs
│   ├── architecture.md       # System architecture deep dive
│   ├── tuning_guide.md       # Miner tuning reference (scoring, parameters, strategy)
│   ├── parameter_ranges.md   # Accepted parameters and valid ranges per tool
│   ├── hap_py_docker.md      # hap.py Docker image reference
│   └── ai-assistant/         # AI assistant knowledge graph, runtime skills, and persona assets
├── scripts/                  # Developer tools
│   ├── verify.sh             # Pre-flight environment check
│   ├── demo.sh               # End-to-end demo runner
│   ├── minosvm_first_run.sh  # Friendly MinosVM first-run launcher
│   ├── prompt_ai_assistant.sh # First-run Ditto/OpenClaw/Hermes prompt
│   ├── setup_ditto_agent.sh  # Optional Ditto @minos graph subscription
│   ├── setup_ai_assistant.sh # Optional Ditto + Minos MCP + OpenClaw/Hermes setup
│   └── auto_update.sh        # PM2 auto-update worker used by enable_auto_update.sh
├── tests/                    # Unit and integration tests
│   ├── conftest.py
│   └── test_*.py             # Tests for scoring, config, platform client, etc.
├── install.sh                # Installer (full setup or update mode)
├── enable_auto_update.sh     # Enable safe git pull + PM2 restart updates
├── disable_auto_update.sh    # Stop the PM2 auto-updater
├── setup.py                  # Interactive setup wizard
├── start-miner.sh            # Start miner (with inline wallet setup)
├── start-validator.sh        # Start validator (with inline wallet setup)
├── pm2-miner.sh              # Start / restart miner under PM2 (wraps start-miner.sh)
├── pm2-validator.sh          # Start / restart validator under PM2 (wraps start-validator.sh)
├── ecosystem.miner.config.js # PM2 app config (miner)
├── ecosystem.validator.config.js # PM2 app config (validator)
├── min_compute.yml           # Minimal compute requirements
├── requirements.txt          # Python dependencies
├── .env.miner.example        # Miner environment configuration
├── .env.validator.example    # Validator environment configuration
└── README.md                 # This document
```

---

## System Prerequisites

| Component | Requirement | Notes |
|-----------|-------------|-------|
| OS | Linux (Ubuntu 20.04+), macOS 13+ | Docker + Bittensor run best on Linux |
| CPU/RAM (Validator) | ≥8 cores / 32 GB RAM | hap.py scoring benefits from cores |
| CPU/RAM (Miner) | ≥4 cores / 8–16 GB RAM | 8 GB for BCFtools, 16 GB for DeepVariant |
| Disk | ≥60 GB (miner) / ≥100 GB (validator) | Reference: ~2 GB miner, ~14 GB validator (SDF expands ~6×). Plus per-round BAMs (~6 GB each) until cleaned. |
| Docker | 20.10+ (24.0+ recommended) | Required for GATK, hap.py, bcftools |
| Python | 3.10+ | We test on 3.12 |
| Bittensor | 10.3.0 (via requirements.txt) | Provides wallet/subtensor/dendrite APIs; 11.x is not supported |

---

## Quick Start

```bash
git clone https://github.com/minos-protocol/minos_subnet.git
cd minos_subnet
bash install.sh          # First-time: full setup (venv, deps, Docker, reference data, wallet)
bash start-miner.sh      # Start as miner (choose one)
# OR
bash start-validator.sh  # Start as validator (choose one)
```

The `start-*.sh` scripts handle wallet setup on first run and automatically repair missing required runtime assets before launch, including reference data and Docker images. Run with `--help` to see all options, or `--setup` to re-run the setup wizard. If you already ran `install.sh` before, running it again will only update dependencies and download any new reference data (use `--fresh` to redo everything).

The installer automatically installs required runtime pieces where supported, including Python packages, Docker on Linux, Node/npm for PM2, missing Docker images, and reference data. It only pauses for choices that should stay under your control, such as wallet/registration, existing `.env` overwrite, and optional AI assistant/runtime setup.

### Try mining without registering

Want to verify your variant-calling pipeline works — and see the score it earns — before bonding TAO or registering a hotkey? Run the miner with **`--demo`**:

```bash
bash start-miner.sh --demo            # ephemeral keypair, no wallet needed
# or equivalently, via the env var:
MINER_DEMO=true bash start-miner.sh   # env var works the same as --flag
```

> `--demo` is a one-shot foreground run — it scores once and exits. Run it directly, not under PM2 (which is for the long-running live miner).

`--demo` is a one-shot onboarding run: it skips the chain entirely, downloads a single fixed, fully-answered sample (BAM + truth), runs your variant caller, and prints the **exact score a validator would compute** — so you confirm your Docker setup, reference data, variant caller, and platform connectivity all work end-to-end *and* see the number your config gets. Nothing is persisted on-chain and no TAO is earned. (The sample is fixed for a repeatable first run; use `--practice` below to score against any sample.)

**Switching from demo to live**: `--demo` needs no wallet and writes nothing. When you're ready to mine for real, run `bash start-miner.sh` (or `bash start-miner.sh --setup`) and follow the wizard to set your wallet and register your hotkey on subnet 107.

### Score your config against any sample (practice mode)

`--demo` scores one fixed sample; **`--practice`** lets you score against **any** of the fully-answered chr19/chr20/chr21/chr22 samples — pick from a menu, run your config, and see the **exact score a validator would compute** — no chain, no wallet, no live round at risk:

```bash
bash start-miner.sh --practice                                   # interactive: pick a sample, see your score
bash start-miner.sh --practice --config configs/gatk.conf        # score a specific config
bash start-miner.sh --practice --config configs/gatk.conf --sample-id <sample-id>   # skip the picker
```

Both `--demo` and `--practice` download truth on purpose so you can iterate offline: compare configs, tune, and only deploy to a live round once you're happy with the score. Neither touches the chain or earns TAO.

**MinosVM:** If using the MinosVM image, everything is pre-installed. Just SSH in and run:

```bash
bash scripts/minosvm_first_run.sh
```

The MinosVM first-run menu can start a demo miner, set up live mining, set up a validator, or launch the Minos Miner AI Assistant flow. You can still run `bash start-miner.sh` or `bash start-validator.sh` directly.

### Optional Minos Miner AI Assistant

Minos includes an optional miner onboarding assistant so new miners can ask Minos-specific questions from a local runtime instead of starting with a blank generic agent. The assistant helps with demo mode, PM2/Docker issues, scoring/eligibility, public endpoint checks, safe paste-back rules, variant-calling concepts, and basic config-tuning.

The general setup menu offers:

1. OpenClaw runtime with Ditto skill, Minos MCP live data, and local Minos skill/persona files. This is the recommended path.
2. Hermes runtime with Ditto skill, Minos MCP live data, and local Minos skill/persona files.
3. Ditto account setup only, subscribed read-only to the public `@minos` knowledge graph.
4. Skip for now.

On MinosVM, Ditto is baseline infrastructure and the first-run prompt only asks whether to enable OpenClaw, enable Hermes, or skip the local runtime.

During a fresh miner install, `bash install.sh` offers to set this up after the setup wizard finishes. You can also subscribe Ditto to the public `@minos` graph later:

```bash
bash start-miner.sh --setup-ditto
# or directly:
bash scripts/setup_ditto_agent.sh
```

If you also want a local agent runtime, use the guided AI assistant setup:

```bash
bash start-miner.sh --setup-ai-assistant
# or directly:
bash scripts/setup_ai_assistant.sh
```

That flow defaults to OpenClaw because it gives miners a local runtime plus the `@minos` graph and Minos MCP live data from `https://mcp.theminos.ai`. If you choose OpenClaw or Hermes from the menu, it installs that runtime's Ditto skill, configures the `minos` MCP connection for current round/leaderboard/miner-history data, and installs a local `minos-miner` skill/persona bundle so the agent understands Minos setup, monitoring, variant calling, public endpoints, and safe tuning even before a Ditto lookup. OpenClaw and Hermes are installed noninteractively by default, so miners stay inside the Minos setup flow. Ditto-only remains available outside MinosVM for miners who only want the public graph subscription and no local runtime yet.

OpenClaw/Hermes still need a model provider. For 1.0.0, Minos does not install or tune local open-source models. During setup, Minos checks whether the selected runtime already has a usable provider; if not, it offers a short provider menu. OpenClaw users can choose OpenAI/Codex, Claude/Anthropic, OpenRouter, OpenClaw guided setup, or skip. Hermes users can choose Hermes/Nous Portal, full Hermes setup, or skip. Provider credentials stay inside OpenClaw, Hermes, or the provider OAuth flow; Minos does not read, log, or store model API keys.

The setup script uses the Ditto CLI to initialize or use your Ditto account, then subscribes it read-only to the public `@minos` knowledge graph. The launch graph is `Minos SN107 Public Knowledge Graph` version `1.0.0`, and subscribed search results should show `subscribed_graph:minos-public-knowledge-graph-1.0.0` with `@minos / Minos SN107 - ...` source context. On MinosVM, the default is public-graph-only: private fallback seeding is disabled unless explicitly overridden. Outside MinosVM, the script can save the same public memory-pack files into that Ditto account as a private fallback copy only if graph subscription/search cannot be verified or fallback seeding is explicitly requested. That fallback is only public Minos docs, never miner secrets, and the preferred source remains the shared `@minos` graph. If you choose OpenClaw or Hermes, setup installs the `heyditto` command globally so runtime skills can read Ditto knowledge directly. It also writes the Ditto API key into the runtime's own env file without printing it: `~/.openclaw/.env` for OpenClaw and `~/.hermes/.env` for Hermes. If you do not already have Ditto CLI auth, it creates a Ditto agent account, prints the private claim URL at the end of setup, and saves the same URL locally with private file permissions. Open that claim URL yourself and do not paste it into public channels, screenshots, or logs before claiming it. The setup does **not** upload your Minos `.env`, wallet files, miner configs, logs, presigned URLs, model API keys, or private validator data.

Minos MCP is configured only for public live subnet tools such as current round, leaderboard, recent rounds, miner history, subnet overview, incentives, and emissions. Ditto explains stable setup and troubleshooting concepts; MCP provides current subnet data.

For OpenClaw on headless/container VMs, Minos runs the foreground gateway under PM2 as `openclaw-gateway` because `systemctl --user` is often unavailable. Check it with `openclaw gateway status` or `pm2 logs openclaw-gateway --lines 80`. Start the local OpenClaw terminal UI with `bash scripts/openclaw-tui.sh`; the wrapper reads the local gateway token without printing it so miners do not hit an extra pairing prompt.

Requirements: Ditto setup needs either `heyditto` or Node/npm for `npx -y @heyditto/cli`. `bash install.sh` installs Node/npm where supported. OpenClaw/Hermes are only needed if you choose a runtime. They still need a model provider configured in that runtime. The recommended assistant path is:

```bash
bash scripts/setup_ai_assistant.sh --with-ditto --openclaw
```

Advanced runtime flags such as `--openclaw`, `--hermes`, `--skip-ditto`, and `--dry-run` are available on `scripts/setup_ai_assistant.sh` directly.

If OpenClaw or Hermes says it cannot use Ditto because `heyditto` is missing or `DITTO_API_KEY` is not configured, rerun the matching setup:

```bash
bash scripts/setup_ai_assistant.sh --with-ditto --openclaw
bash scripts/setup_ai_assistant.sh --with-ditto --hermes
```

After setup, try asking Ditto:

```text
PM2 says my Minos miner is online but I have 0 weight. What should I check?
```

In normal operation, agents subscribe to the shared public `@minos` graph. Private fallback seeding is only for cases where graph retrieval cannot be verified or the operator explicitly requests a local fallback copy.

### Running with PM2 (optional)

[PM2](https://pm2.keymetrics.io/) keeps the miner or validator running with restarts and log management. This repo runs the same **`start-miner.sh`** / **`start-validator.sh`** entrypoints under PM2 (so your venv, `.env`, and prerequisite checks stay identical to a manual start).

1. **Install PM2** — **`bash install.sh`** installs or repairs Node.js/npm, then runs **`npm install -g pm2`** automatically. If PM2 installs under an npm global directory that is not on `PATH`, the installer prints the exact `export PATH=...` line to add.
2. **Create `.env` first** — Run **`bash start-validator.sh`** or **`bash start-miner.sh`** once in a normal terminal so wallet/setup can run (PM2 does not provide a TTY for the interactive wizard).
3. **Launch under PM2:**

```bash
bash pm2-validator.sh   # validator: start, or restart if already registered
bash pm2-miner.sh       # miner: start, or restart if already registered
```

Or start from the ecosystem files directly:

```bash
pm2 start ecosystem.validator.config.js
pm2 start ecosystem.miner.config.js
```

Useful commands: **`pm2 status`**, **`pm2 logs minos-validator`** / **`pm2 logs minos-miner`**, **`pm2 save`** (persist process list after reboot — pair with **`pm2 startup`** per PM2 docs). The interactive setup wizard auto-generates **`ecosystem.<role>.config.js`** when PM2 is available. Demo or registered nodes are started automatically; unregistered live nodes are registered in PM2 but left stopped until you register the hotkey.

### Automatic Updates with PM2

If your miner or validator already runs under PM2, you can enable safe automatic git pulls and PM2 restarts:

```bash
./enable_auto_update.sh
```

The setup script finds likely PM2 miner/validator processes, asks you to confirm the exact process, then starts a small PM2 updater named **`minos-auto-update`**.

Safety rules:

- Local git changes or untracked files make the updater skip the pull.
- Only fast-forward git updates are applied.
- Only the PM2 process you confirmed is restarted.
- The selected repo, branch, and PM2 process are saved in `~/.minos/auto_update.env`.

Useful commands:

```bash
pm2 logs minos-auto-update
./disable_auto_update.sh
```

### Updating

Already running Minos? Pull the latest code and run the installer — it detects your existing setup and only downloads new reference data:

```bash
git pull
bash install.sh
```

Your wallet, `.env`, and existing data are preserved. Use `bash install.sh --fresh` to redo a full setup if needed.

<details>
<summary>Manual setup (if you prefer not to use the install script)</summary>

#### 1. Setup Environment

```bash
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```

#### 2. Configure Environment

```bash
cp .env.miner.example .env    # for miners
cp .env.validator.example .env # for validators
```

#### 3. Pull Docker Images

```bash
docker pull broadinstitute/gatk:4.5.0.0
docker pull google/deepvariant:1.5.0
docker pull genonet/hap-py@sha256:03acabe84bbfba35f5a7234129d524c563f5657e1f21150a2ea2797f8e6d05f2
docker pull quay.io/biocontainers/bcftools:1.20--h8b25389_0
docker pull quay.io/biocontainers/samtools:1.20--h50ea8bc_0
# Validators only — the freebayes image is retained until in-flight
# pre-cutover rounds finish scoring, then removed in a follow-up release:
# docker pull staphb/freebayes:1.3.7
```

> **Note:** The hap.py image is pinned by SHA256 digest for reproducible scoring. The tag `genonet/hap-py:0.3.15` points to the same image, but validators use the pinned digest for consistency.

#### 4. Run Setup Wizard

```bash
# Interactive wizard: configures wallet, downloads reference data, sets up .env
python setup.py
```

</details>

---

## Validator Setup

### Environment Variables

```bash
# Network
NETUID=107

# Wallet
WALLET_NAME=validator
WALLET_HOTKEY=default

# Platform
PLATFORM_URL=https://api.theminos.ai
PLATFORM_TIMEOUT=60

# Storage preference: hippius = Hippius SN75 first (recommended, Bittensor decentralized).
# aws_s3 = R2/AWS first. Both serve the same files; this just controls fetch order.
STORAGE_PRIMARY_BACKEND=hippius
```

### Running the Validator

```bash
bash start-validator.sh
bash start-validator.sh --wallet-name validator   # Pre-fill wallet name
bash start-validator.sh --setup                   # Re-run setup wizard
```

For production-style supervision, use **[Running with PM2 (optional)](#running-with-pm2-optional)** (`bash pm2-validator.sh`).

Or manually:

```bash
source .venv/bin/activate
python -m neurons.validator \
  --netuid 107 \
  --subtensor.network finney \
  --wallet.name validator \
  --wallet.hotkey default
```

### Validator Workflow

1. **Fetch Rounds and Assignment**: Poll platform for scoring rounds and this validator's primary/secondary miner assignment
2. **Download Round Data**: Download the mutated BAM, merged truth VCF, and mutations-only VCF from platform presigned URLs
3. **Run Miner Tools**: Execute assigned miners' variant-calling configs via templates, scoring primaries first and secondaries only while enough time remains
4. **Scoring**: Validate each generated VCF with hap.py against the platform truth data, then compute the AdvancedScorer result
5. **Submit Scores**: Report each per-miner score and audit artifact pointers back to the platform
6. **Backfill and Record Round**: After the scoring window closes, fetch peer scores for miners not personally covered and record complete participation
7. **Weight Update**: Submit weight history to the platform, then set on-chain weights if the validator is registered

### Validator Parallelism

Each round, the validator runs many miners' variant-calling configs concurrently. Concurrency, per-job thread count, and memory ceiling are auto-tuned at startup from host CPU/RAM:

- Reserves 4 cores + 16 GB for the OS, Docker daemon, and hap.py
- Picks `threads_per_job` between 2 and 8 (DeepVariant's call_variants is single-threaded past 8)
- Pins memory at 16 GB per job (DeepVariant's documented minimum)
- `concurrency = min(usable_cores // threads_per_job, usable_ram // 16, 8)`

Examples:

| Host | Auto-tune |
|---|---|
| 8c / 32 GB | 1 concurrent × 2 threads × 16 GB |
| 16c / 64 GB | 3 concurrent × 3 threads × 16 GB |
| 32c / 128 GB | 4 concurrent × 7 threads × 16 GB |
| 64c / 256 GB | 7 concurrent × 8 threads × 16 GB |

To override, set `MINOS_VALIDATOR_CONCURRENCY`, `SCORING_THREADS`, or `SCORING_MEMORY_GB` in `.env`. Note `SCORING_MEMORY_GB < 16` will OOM-crash DeepVariant submissions.

---

## Miner Setup

### Environment Variables

```bash
# Network
NETUID=107

# Wallet
WALLET_NAME=miner
WALLET_HOTKEY=default

# Miner
MINER_TEMPLATE=gatk

# Platform
PLATFORM_URL=https://api.theminos.ai
PLATFORM_TIMEOUT=60

### Running the Miner

```bash
bash start-miner.sh
bash start-miner.sh --wallet-name miner --miner-template deepvariant  # Pre-fill values
bash start-miner.sh --setup                                           # Re-run setup wizard
```

For production-style supervision, use **[Running with PM2 (optional)](#running-with-pm2-optional)** (`bash pm2-miner.sh`).

Or manually:

```bash
source .venv/bin/activate
python -m neurons.miner \
  --netuid 107 \
  --subtensor.network finney \
  --wallet.name miner \
  --wallet.hotkey default
```

### Miner Workflow

1. **Registration**: Register with platform via hotkey authentication
2. **Task Poll**: Poll platform for pending evaluation tasks
3. **BAM Download**: Fetch benchmark BAM from platform via presigned URL
4. **Variant Calling**: Run configured variant caller (GATK, DeepVariant, or bcftools — freebayes deprecated 2026-05-09)
5. **Config Submit**: Submit tool config you've used (hyperparameters only based on the template)
6. **Reward**: Earn alpha based on accuracy score — validators re-run the config to verify

---

## Platform Service

The Minos Platform is a hosted service at `https://api.theminos.ai` that handles BAM generation and synthetic mutation injection with HelixForge, file transfers via presigned URLs, and round coordination. Validators and miners connect automatically via the `PLATFORM_URL` environment variable — no self-hosting required.

### Key Endpoints

| Endpoint                         | Used by   | Description                                    |
| ---------------------------------|-----------|------------------------------------------------|
| `POST /v2/round-status`         | Miner     | Poll for active rounds and presigned BAM URL   |
| `POST /v2/submit-config`        | Miner     | Submit variant-calling tool config             |
| `POST /v2/get-scoring-rounds`   | Validator | Fetch rounds ready for scoring                 |
| `POST /v2/get-submissions`      | Validator | Fetch all miner configs for a round            |
| `POST /v2/get-assignment`       | Validator | Get primary/secondary miner scoring assignment |
| `POST /v2/submit-score`         | Validator | Submit per-miner scores                        |
| `POST /v2/get-backfill-scores`  | Validator | Fetch peer scores after scoring window closes  |
| `POST /v2/submit-weight-history`| Validator | Submit round scores, eligibility, and weights   |
| `POST /v2/get-validator-state`  | Validator | Recover round score state after validator restart |

### Storage Backends

The platform serves all per-round files (BAMs, truth VCFs, mutations VCFs) via short-lived presigned URLs and a primary + backup pair, so a miner or validator never needs storage credentials of its own. Three backends sit behind the platform:

| Backend          | Role                                                                          |
|------------------|-------------------------------------------------------------------------------|
| Cloudflare R2    | Primary for per-round artifacts when enabled platform-side (free egress)      |
| AWS S3           | Fallback for per-round artifacts (`genotypenet-platform` bucket)              |
| Hippius (SN75)   | Decentralized backup, Bittensor-native (`genotypenet-mutations` bucket)       |

Reference data (FASTA, FAI, dict, RTG SDF) is served via the indirected URL `https://api.theminos.ai/reference/...` which 302-redirects to the active backend (currently a public R2 bucket). Setup downloads through this URL only — no direct R2/S3 hostnames are baked into the subnet code.

The `STORAGE_PRIMARY_BACKEND` env var only controls **fetch order on the client side**:

- `hippius` (default) — try the Hippius URL first, fall back to the R2/AWS URL.
- `aws_s3` — try the R2/AWS URL first, fall back to Hippius. (The label says "aws_s3" for backwards compatibility, but the platform may serve either R2 or AWS in this slot — both behave identically from the client's perspective.)

Both URLs always point at the same files. Pick whichever is faster from your network.

---

## Scoring System

### hap.py Validation

Validators run each miner's tool config and score the resulting VCF from that against the truth data shared by the platform with them using hap.py. Scores are combined by Minos' developed `AdvancedScorer` into a scaled 0–100 final score that balances accuracy and precision with the following components:

| Component    | Weight |
| -------------|--------|
| Core F1      | 60%    |
| Completeness | 15%    |
| FP Rate      | 15%    |
| Quality      | 10%    |

### Round Score Tracking

The raw AdvancedScorer output (0-100) is normalized to a 0-1 scale and used as the miner's score for that round. Winners are selected from the current round's valid Advanced Score only; historical score carryover is not used.

```python
# AdvancedScorer returns 0-100, normalized to 0-1 for round scoring
combined_final = advanced_score / 100.0
current_score = combined_final
```

Only finite scores where `0 < current_score <= 1.0` are eligible for ranking. Rounds with no valid local or backfilled scores fail closed and do not reuse previous-round scores. The validator and platform also filter the synthetic all-zero input fingerprint around 0.25 so empty or invalid calls cannot become winners.

### Weight Distribution

Validators apply the reward policy from `/scoring/network-config` (winner-heavy pruning dust). Currently the top eligible current-round miner receives ~90% of the weight (`winner_weight = 0.9`), eligible ranks #2 through #20 split the remaining ~10% as ranked pruning dust (`dust_top_n = 20`) using 0.8 geometric weighting, and burn is 0% (`burn_rate = 0.0`). These are dynamic protocol values — **check `/scoring/network-config` for the latest**. If fewer dust recipients exist, the dust is renormalized across the available ranks; any weight that cannot be assigned to an eligible miner is sent to burn.

Eligibility requires scoring in at least 5 of the last 20 rounds, including the current round. Miners below the participation threshold receive 0 weight until they qualify. Close current-round ties use deterministic round data from the validator/platform flow; historical scores are not part of the ranking.

---

## Monitoring & Troubleshooting

### Common Issues

| Symptom                       | Cause                                      | Fix                                                                                                              |
|-------------------------------|--------------------------------------------|------------------------------------------------------------------------------------------------------------------|
| `docker: permission denied`   | User not in docker group                   | `sudo usermod -aG docker $USER && newgrp docker`                                                                 |
| `GATK timeout`                | Insufficient resources                     | Increase threads/memory or timeout                                                                               |
| DeepVariant OOM / killed      | <16 GB available to the container          | DeepVariant needs ≥16 GB. Free RAM, close other tools, or switch template to GATK/BCFtools                       |
| `Platform 401 error`          | Invalid sig or unregistered hotkey         | Ensure wallet hotkey is registered on the metagraph                                                              |
| `No miners available`         | No registered miners                       | Check metagraph for active miners                                                                                |
| `hap.py zero scores`          | VCF format issues                          | Ensure single-sample VCF output                                                                                  |
| Round skipped, "<10min left"  | Submitted too late                         | Miner needs ≥10 min remaining to start a round. Check clock skew (`timedatectl`) and platform connectivity speed |
| Reference download stuck/slow | Platform redirect or transient network     | Setup retries once automatically. If it still fails, re-run `python setup.py --update-data-only`                 |
| Validator: "no scoring rounds"| Round still in submission window           | Validators only score AFTER the submission window closes. Wait for the next tempo boundary                       |
| Earning 0 weight as miner     | Not eligible yet, no valid current score, or outside the top current-round winners | Need >=5 of last 20 rounds participated, then a stronger current-round Advanced Score. See [tuning_guide](docs/tuning_guide.md) |

### Logs

```bash
# PM2 (recommended)
pm2 logs minos-miner
pm2 logs minos-validator

# systemd (if using systemd service)
journalctl -u minos-miner -f
journalctl -u minos-validator -f

# Direct (if running manually)
# Output streams to terminal
```

### Health Checks

```bash
# Platform health
curl https://api.theminos.ai/health

# Metagraph status
btcli subnet metagraph --netuid 107
```

---

## Additional Documentation

- [neurons/README.md](neurons/README.md) - Detailed miner/validator documentation
- [utils/README.md](utils/README.md) - Utility modules reference
- [docs/architecture.md](docs/architecture.md) - System architecture deep dive
- [docs/tuning_guide.md](docs/tuning_guide.md) - Miner tuning guide (scoring breakdown, parameters, strategy)
- [docs/parameter_ranges.md](docs/parameter_ranges.md) - Accepted parameters and valid ranges per tool (mirrors `GET /scoring/parameter-ranges`)
- [docs/hap_py_docker.md](docs/hap_py_docker.md) - hap.py Docker image reference
- [docs/ai-assistant/memory-pack/README.md](docs/ai-assistant/memory-pack/README.md) - Public Minos SN107 knowledge source for Ditto and agent runtimes
- [docs/ai-assistant/README.md](docs/ai-assistant/README.md) - OpenClaw/Hermes local skill, persona, and model setup assets
- [scripts/verify.sh](scripts/verify.sh) - Pre-flight environment check (`bash scripts/verify.sh --miner`)
- [scripts/demo.sh](scripts/demo.sh) - Run a single demo round end-to-end (`bash scripts/demo.sh`)
- [scripts/minosvm_first_run.sh](scripts/minosvm_first_run.sh) - Friendly MinosVM first-run menu
- [scripts/prompt_ai_assistant.sh](scripts/prompt_ai_assistant.sh) - One-time AI assistant prompt for installs and VM first run
- [scripts/setup_ditto_agent.sh](scripts/setup_ditto_agent.sh) - Subscribe Ditto to the public `@minos` knowledge graph
- [scripts/setup_ai_assistant.sh](scripts/setup_ai_assistant.sh) - Set up Ditto first, then optionally connect OpenClaw or Hermes

---

## Links

- **GitHub**: [github.com/minos-protocol/minos_subnet](https://github.com/minos-protocol/minos_subnet)

minos_subnet

minos_subnet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

minos_subnet

minos_subnet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README