jebat-core

# JEBAT v8.2.1 — Sovereign Agent OS & Agent Workstation        > **Sovereign execution, private memory, and audited intelligence.** JEBAT is an enterprise-grade self-hosted AI platform and agent workstation. It provides governed local LLM inference, multi-agent orchestration, persistent memory, adaptive context windowing, observability, an MCP workspace server, and a full web interface. Run fully air-gapped on your private network with zero data leakage. Named after the legendary Malay warrior **Hang Jebat** — loyal, powerful, and unforgettable. ## Public Service The public JEBAT service is available at [`jebat.online`](https://jebat.online). | Surface | URL | |---------|-----| | API status | [`/api/status`](https://jebat.online/api/status) | | Swagger UI | [`/swagger`](https://jebat.online/swagger) | | OpenAPI document | [`/openapi.json`](https://jebat.online/openapi.json) | | WebUI | [`/webui/`](https://jebat.online/webui/) | The public API and WebUI are production services. Configure provider credentials in your own deployment before sending requests that require model inference. --- ## Quick Start ### Option 1: npx (Recommended — Zero Install) ```bash # Interactive REPL npx @nusabyte/jebat repl # One-shot chat npx @nusabyte/jebat chat "Explain the memory consolidation algorithm" # Run agent task npx @nusabyte/jebat agent "Audit all API endpoints in src/services" # Generate code npx @nusabyte/jebat code "Create a REST API with FastAPI" # Launch WebUI npx @nusabyte/jebat webui # With bun bunx @nusabyte/jebat repl ``` **Requirements:** Node.js >= 18, Python >= 3.11, pip ### Option 2: Global Install (via npm) ```bash # Installs the npx wrapper globally; first run bootstraps the Python core npm install -g @nusabyte/jebat jebat repl ``` > The npm package is a thin launcher — on first run it downloads and runs the > bootstrap, which provisions Python, the venv, and the `jebat` launcher in > `~/.local`. **Do not run `pip install jebat`** — that name is a unrelated > package on PyPI (a website checker) and is **not** JEBAT. The bootstrap is the > only supported install path. ### Option 3: Source Build ```bash git clone https://github.com/nusabyte-my/jebat-core.git cd jebat-core curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash # bootstrap the core CLI jebat repl ``` --- ## Installation — Three Surfaces, One Script JEBAT installs like Hermes: a single self-contained bootstrap that provisions Python, the venv, the CLI launcher, and (optionally) the Desktop app and MCP server. **No sudo required** — everything lands in `~/.local`. ```bash # Zero-install (downloads + runs the bootstrap). Uses the raw GitHub URL so it # always serves the script as plaintext (no CDN/HTML fallback). curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash # CLI only curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --desktop curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --mcp curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --all # Or pin a profile explicitly curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --cli --no-inference # Agent / CI friendly: fully non-interactive curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --yes --quiet # Emit a ready-to-paste MCP config for your IDE (no install needed) curl -fsSL https://raw.githubusercontent.com/nusabyte-my/jebat-core/main/install.sh | bash -s -- --print-mcp-config=cursor ``` ### What gets installed (components) | Component | Surface | Purpose | |-----------|---------|---------| | **JEBAT CLI** | CLI / Desktop / MCP | `jebat repl`, `chat`, `code`, `agent` (core) | | **Python venv** | all | Isolated runtime at `~/.local/jebat/venv` (editable install) | | **`jebat` launcher** | all | Symlink on `PATH` (`~/.local/bin/jebat`) | | **DESIGNmd CLI** | CLI companion | Standalone registry tool (`designmd search/get/download/...`) | | **MCP surface** | MCP | IDE snippet prepped at `~/.local/jebat/mcp.ide.json` (server ships with the full Jebat workspace) | | **Unified Inference Stack** | CLI / Desktop | Ollama + llama.cpp + router + OpenWebUI (from the full Jebat workspace) | | **Desktop surface** | Desktop | `jebat desktop` UI from the full Jebat workspace | > **Note:** This installer clones the public **core** CLI (`jebat-core`), > which exposes `repl`, `chat`, `code`, and `agent`. The `design`, `mcp serve`, > and `desktop` surfaces live in the full **Jebat workspace** (`jebat-dev`). > The installer still prepares their config (IDE snippet, Electron note) so > they activate the moment that code is present. ### Surface quick-reference ```bash # CLI — primary surface (core) jebat repl # interactive REPL jebat chat "hello" # one-shot jebat code "refactor utils.py" # coding agent jebat --help # full command list # Default inference provider is seeded at install: JEBAT Online # (OpenAI-compatible) → https://jebat.online/v1 [switch with /provider <id>] # local Ollama (http://127.0.0.1:11434) is also registered, inactive by default # DESIGNmd (companion tool, standalone) designmd search "dark fintech" --limit 3 # browse the registry designmd tags # list tags # MCP — run from a full workspace checkout # python ./jebat-mcp --transport stdio # See MCP.md for IDE configuration and self-hosted HTTP deployment. # Desktop — native workstation UI (full workspace) jebat desktop # launch GUI ``` ### Local Unified Inference Stack (optional, `--no-inference` to skip) The full Jebat workspace ships a local LLM layer so `jebat` can route to your own hardware: - **llama.cpp** (`llama-server`) — CPU/AVX-512 + Intel Iris Xe (Vulkan) - **Ollama** — local GGUF registry - **Router** — one OpenAI-compatible endpoint at `http://127.0.0.1:8000/v1` that merges every backend and fans out by model/`@backend`/`?backend=` - **OpenWebUI** — chat frontend at `http://127.0.0.1:8080`, pointed at the router - **Remote vLLM / Ollama** — offload big models to a GPU host ```bash # from the full Jebat workspace checkout: infra/inference/start.sh # boot the stack # or as a user service: systemctl --user start jebat-router.service ``` --- ## Remote Ollama Endpoint JEBAT uses a remote Ollama instance hosted on a dedicated server with AMD EPYC 9354P (8 vCPU, 32GB RAM). ```bash # Default endpoint (zero config) # https://jebat.online/ollama # Models available: # - qwen2.5-coder:7b (4.7GB, fast coding) # - qwen2.5:14b (9GB, general purpose) # Test the endpoint curl https://jebat.online/ollama/v1/models # Use with JEBAT jebat chat "Hello from jebat.online" ``` ### Configuration ```yaml # ~/.jebat/config.yaml llm: provider: ollama model: qwen2.5-coder:7b api_base: https://jebat.online/ollama # Default endpoint ``` --- ## Providers & Auth JEBAT routes to **17 providers** — OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, Cerebras, Together, OpenRouter, GitHub Models, Cloudflare, SambaNova, Novita, Z.AI, local Ollama, or any OpenAI-compatible endpoint. Add one interactively from the REPL: ```bash jebat repl # then: /provider add ``` ### Auth method — `key` / `env` / `store` Every keyed provider (Gemini, Anthropic, and any OpenAI-compatible) supports three ways to supply the API key. Choose at add time; the key is resolved at runtime via `resolve_api_key()`. | Method | How it works | Best for | |--------|--------------|----------| | `key` | Paste the key inline (stored in `~/.jebat/jebat-cli-providers.json`) | Quick local use | | `env` | Read from an env var at runtime (e.g. `ANTHROPIC_API_KEY`) | CI / shared machines — secret never hits disk | | `store`| Named key from the JEBAT auth store (`~/.jebat/auth/tokens.json`, set via `/apikey`) | Multiple / rotated keys | `env` and `store` keep secrets out of the provider config — recommended for anything shared. ### Models (refreshed catalog) Defaults point at current models: - **OpenAI** → `gpt-5` (+ `gpt-5-mini` / `gpt-5-nano`), `gpt-4.1`, `o3`, `o4-mini` - **Anthropic** → `claude-opus-4-1` / `claude-sonnet-4-1` / `claude-haiku-4-5` - **xAI** → `grok-4` (+ `grok-4-mini`) - **Gemini** → `gemini-2.5-pro` / `gemini-2.5-flash` - **GitHub Models** → `openai/gpt-5` (free tier with a GitHub token) - **Local Ollama** → `qwen2.5:3b` (avoids a 404 when a bigger model isn't pulled) ### Local Ollama auto-start When Ollama is the selected provider, JEBAT auto-starts `ollama serve` if it isn't running and auto-pulls the chosen model if it isn't pulled — no manual setup. ### Default provider Install seeds **JEBAT Online** (`https://jebat.online/v1`, OpenAI-compatible) as the active default; local Ollama is registered but inactive. Switch any time with `/provider <id>`. --- ## Platform Suite — 5 Products, 1 Memory JEBAT ships as a unified platform with 5 integrated products sharing a persistent memory layer and MCP protocol bus. | Product | Codename | Description | |---------|----------|-------------| | **JEBAT Core** | — | The sovereign AI engine — REPL, chat, agent, memory, MCP | | **Sahabat** | Companion | Memory keeper, briefing agent, meeting intelligence | | **Keris** | Sentinel | Autonomous pentest orchestrator with AI analysis | | **Pandai** | Developer | Skills, tools, code generation, developer suite | | **Perisai** | Nexus | Multi-channel bot orchestrator (Telegram, Discord, Slack, etc.) | --- ## WebUI — Stealth-Dark Tactical Interface Launch the full web interface with real-time system metrics, product dashboards, and enterprise UX. ```bash # Launch WebUI python -m jebat.services.webui.launch # → http://localhost:8787 # Or via CLI jebat webui ``` ### WebUI Pages (13) | Page | Description | |------|-------------| | **Dashboard** | System metrics (CPU/RAM/disk), product overview, recent scans | | **Chat** | Universal AI chat with provider selection | | **Sahabat** | Companion chat, drag-and-drop transcript upload | | **Keris** | Security scan form, scan history, severity badges | | **Pandai** | Skills browser, code tools | | **Perisai** | Channel management, send/broadcast, channel stats | | **Control** | Runtime control, provider config | | **Skills** | Skill browser and management | | **Agents** | Agent orchestration and status | | **Channels** | Messaging channel configuration | | **Setup** | Environment setup wizard | | **Settings** | Provider selection, preferences, API reference | ### WebUI Enterprise Features (25) **Backend:** - Rate limiting (120 req/min per IP) - CORS tightened (localhost-only) - CSP headers (Content-Security-Policy, X-Frame-Options, nosniff) - Request ID tracking (X-Request-ID on every response) - Audit trail (last 500 requests, `/api/audit`) - Error pages (404/429/500 styled) - Health check (`/health`) - Session check (`/api/session/check`) **Frontend:** - Loading progress bar - Dark/Light theme toggle - Clickable breadcrumb navigation - Global search palette (Ctrl+K) - Keyboard shortcuts overlay (?) - Toast notification system - Activity log drawer (Ctrl+L) - Activity log export (JSON/CSV) - Confirmation dialogs - Session timeout (15 min inactivity) - i18n (English/Bahasa Melayu) - Copy-to-clipboard - Focus trap (keyboard accessibility) - Skeleton loading (shimmer) - Reduced motion support (prefers-reduced-motion) - Print-friendly CSS - Drag-and-drop file upload - Service worker (offline caching) - PWA manifest (install as desktop app) - Favicon + OG meta tags - Responsive mobile sidebar ### WebUI API Endpoints ``` GET /health Health check GET /api/system/metrics CPU, memory, disk, uptime GET /api/audit Audit trail (last N requests) GET /api/session/check Session verification GET /api/keris/history Scan history POST /api/keris/scan Trigger pentest scan GET /api/nexus/channels List channels POST /api/nexus/send Send message POST /api/nexus/broadcast Broadcast to all GET /api/nexus/stats Channel statistics GET /webui/api/status Provider/model status POST /webui/api/chat Chat with AI POST /webui/api/runtime Runtime control ``` --- ## Core Commands ### General | Command | Description | |---------|-------------| | `jebat repl` | **Interactive REPL** — streaming, tools, history | | `jebat chat "prompt"` | One-shot chat with tool calling | | `jebat agent "task"` | Run one-shot agent with tool-calling | | `jebat code "prompt"` | Generate code from description | | `jebat webui` | Launch Stealth-Dark WebUI | | `jebat status` | System health & provider status | | `jebat doctor` | Diagnose environment issues | | `jebat init` | Initialize JEBAT with provider config | ### Configuration & Memory | Command | Description | |---------|-------------| | `jebat config show\|set\|reset\|edit` | Configuration management | | `jebat memory store\|search\|stats` | 6-type eternal memory with Ghost DB vector search | | `jebat llm providers\|config\|auth` | LLM provider management | | `jebat llm best-provider` | Auto-detect best available provider | ### File & Tools | Command | Description | |---------|-------------| | `jebat file read\|write\|patch\|search\|undo\|tree` | Safe file ops with backups | | `jebat tools list\|inspect` | Inspect all registered tools | | `jebat skills list\|search\|show` | Tok Guru skills | ### Platform Suite | Command | Description | |---------|-------------| | `jebat companion chat\|briefing\|meeting\|stats` | Sahabat — companion AI | | `jebat keris scan\|assess\|history` | Keris — pentest scanner | | `jebat nexus list\|add\|remove\|send\|broadcast\|health\|stats` | Perisai — multi-channel bot | | `jebat design search "query" [--tag X] [--limit N] [--json]` | Search the DESIGNmd.ai registry (no key needed) | | `jebat design get <owner/name> [--json]` | View a design system (needs DESIGNMD_API_KEY) | | `jebat design download <owner/name> [-o PATH \| --vault]` | Save a system as DESIGN.md (needs key) | | `jebat design upload ./DESIGN.md --name "Kit" --tags a,b` | Publish a DESIGN.md to your account (needs key) | | `jebat design tags` | List registry tags (no key needed) | | `jebat design sync --tag dark --limit 5 [--vault]` | Bulk-pull tag-matched systems into the vault | | `jebat design doctor` | Check install / API key / connectivity | ### Agent Orchestration | Command | Description | |---------|-------------| | `jebat delegate run "task"` | Multi-agent swarm dispatch | | `jebat loop start\|stop\|status` | Autonomous agent loop | | `jebat think "question"` | Deep reasoning engine | --- ## MCP Server Integration JEBAT includes a native MCP server for a full workspace checkout. Use the standalone `jebat-mcp` entry point for local IDE connections. The available tool set is determined by the installed runtime and its enabled integrations; do not expose it publicly without access controls. ### Running JEBAT as an MCP Server ```bash # Stdio mode (recommended for a local IDE) python ./jebat-mcp --transport stdio # Self-hosted HTTP mode (bind to loopback and put authentication in front of it) python ./jebat-mcp --transport http --host 127.0.0.1 --port 8099 ``` ### MCP Tools JEBAT exposes registered file, terminal, browser, search, scheduling, memory, and integration tools over MCP. Tool availability depends on the installed workspace and enabled integrations. | Category | Tools | |----------|-------| | **File** | `file_read`, `file_write`, `file_patch`, `file_search`, `file_undo`, `file_tree` | | **Terminal** | `terminal`, `terminal_bg`, `process_list`, `process_log`, `process_kill`, `process_write` | | **Browser** | `browser_navigate`, `browser_snapshot`, `browser_click`, `browser_type`, `browser_scroll`, `browser_back`, `browser_vision`, `browser_console`, `browser_get_images` | | **Vision / Search** | `vision_analyze`, `image_generate`, `search_web`, `web_extract` | | **Auth / Cron** | `auth_*`, `cron_*` (create, list, pause, resume, remove, update, run) | | **Wiki** | `wiki_*` (create, read, edit, delete, list, search, auto_save, suggest, consolidate, stats) | For a remote installation, deploy the MCP service with `infra/vps/vps/docker-compose.mcp.yml` and an authenticated reverse proxy based on `infra/vps/vps/nginx.jebat.mcp.conf`. The repository does not promise a shared public MCP endpoint. ### IDE Configuration #### Cursor / Windsurf / Zed Add to your MCP config (`.cursor/mcp.json`, `.windsurf/mcp.json`): ```json { "mcpServers": { "jebat": { "command": "python3", "args": ["/absolute/path/to/jebat-mcp"], "env": {} } } } ``` #### VS Code (via Continue / Cline extension) ```json { "mcpServers": { "jebat": { "command": "python3", "args": ["/absolute/path/to/jebat-mcp"] } } } ``` #### JetBrains (via MCP Plugin) Settings > Tools > MCP > Add Server: ``` Command: python3 /absolute/path/to/jebat-mcp Mode: stdio ``` ### Connecting to External MCP Servers ```bash # Start a local MCP server for your IDE python ./jebat-mcp --transport stdio # Start a loopback HTTP service behind your own authenticated proxy python ./jebat-mcp --transport http --host 127.0.0.1 --port 8099 ``` --- ## Keris — Sentinel Security Autonomous pentest orchestrator with AI-powered analysis. ```bash # Quick scan jebat keris scan example.com --profile quick # Full scan with orchestration jebat keris scan 192.168.1.0/24 --profile full # Quick assessment jebat keris assess example.com # View history jebat keris history ``` ### Scan Profiles | Profile | Description | |---------|-------------| | `quick` | Port scan + basic vulns (~30s) | | `standard` | Full port scan + service detection + vulns (~2min) | | `full` | Comprehensive scan + CVE check + SSL analysis (~5min) | | `vuln` | Vulnerability-focused scan only | --- ## Perisai — Nexus Multi-Channel Multi-channel bot orchestrator for Telegram, Discord, Slack, Signal, Matrix, and WhatsApp. ```bash # List channels jebat nexus list # Add a channel jebat nexus add telegram -123456789 --token "BOT_TOKEN" # Send a message jebat nexus send -123456789 "Hello from JEBAT!" # Broadcast to all channels jebat nexus broadcast "System update complete" # Health check jebat nexus health # Statistics jebat nexus stats ``` --- ## Sahabat — Companion AI companion for memory, briefings, and meeting intelligence. ```bash # Interactive chat jebat companion chat # Generate briefing jebat companion briefing # Summarize meeting jebat companion meeting --file transcript.txt --title "Sprint Review" # Companion stats jebat companion stats ``` --- ## Multi-Agent Orchestration JEBAT dispatches work to specialized sub-agents: ```bash # Delegate to a specific agent jebat delegate run "Audit all API endpoints" --agent tukang --tools terminal,file # Parallel multi-agent task jebat delegate run "Security audit + code review" --parallel # Spawn agent swarms jebat agent "Analyze this codebase for vulnerabilities, performance issues, and code quality" ``` ### Agent Roles | Agent | Domain | |-------|--------| | **Tukang** | Implementation, code, builds | | **Hulubalang** | Security, pentest, hardening | | **Pawang** | Research, investigation, docs | | **Syahbandar** | Ops, automation, deploy | | **Bendahara** | Database, schema, migrations | | **Penyemak** | QA, validation, release | --- ## Technical Comparison | Capability | JEBAT v8.2.1 | Commercial SaaS (Claude/GPT) | Ollama WebUI | LM Studio | | :--- | :---: | :---: | :---: | :---: | | **Data Residency** | **100% Private / Air-gapped** | Cloud (Third-Party) | Local Only | Local Only | | **LLM Provider Routing** | **17 Providers (Failover)** | Single Provider | Ollama Only | Local Only | | **MCP Server** | **Native workspace server** | None | Basic API | Local API | | **Cognitive Profiles** | **7 Thinking Modes (Ultra-Think)** | Standard Chat | Standard Chat | Standard Chat | | **Security Auditing** | **Autonomous Pentest Suite** | Blocked | None | None | | **Access Control (RBAC)** | **3-Tier Command Classification** | Muted Policy | Basic Auth | None | | **Eternal Memory** | **5-Layer Heat-Scored Recall** | Session Limits | No | No | | **Web Interface** | **Stealth-Dark Tactical (25 features)** | Chat UI | Basic UI | Local UI | | **Platform Suite** | **5 Products, 1 Memory** | Single Product | None | None | | **npx / npm** | **Zero-install wrapper** | N/A | N/A | N/A | | **Remote Inference** | **AMD EPYC 9354P (8 vCPU/32GB)** | Cloud GPU | Local GPU | Local GPU | --- ## Configuration Config file: `~/.jebat/config.yaml` ```yaml llm: provider: ollama model: qwen2.5-coder:7b api_base: https://jebat.online/ollama fallback_providers: - openai - anthropic - openrouter - groq safety: default_tier: confirm sandbox_restricted: true mcp: server_port: 8787 enable_stdio: true agent: max_iterations: 10 stream_tokens: true ``` ### Quick Config Commands ```bash # Show config jebat config show # Set value jebat config set agent.safety_mode confirm # Edit in $EDITOR jebat config edit # Health check jebat doctor ``` --- ## System Architecture ``` jebat-core/ ├── jebat/ # Active runtime modules │ ├── cli/ # Entrypoints │ ├── core/ # Cognitive loops (agent_loop, orchestrator, delegation) │ ├── mcp/ # MCP server + skill registry + adapter │ ├── services/ # WebUI, MCP protocol, API gateway │ │ └── webui/ # Stealth-Dark Tactical WebUI │ │ ├── launch.py # Enterprise launcher (rate limit, CSP, audit) │ │ ├── webui_server.py # FastAPI server + WebSocket │ │ └── static/ # SPA shell, 13 partials, CSS theme │ ├── features/ # Specialized capability suites │ │ ├── companion/ # Sahabat — memory, briefing, meetings │ │ ├── sentinel/ # Keris — pentest orchestrator │ │ ├── nexus/ # Perisai — multi-channel bot │ │ ├── design/ # DESIGN.md integration │ │ ├── code_agent/ # Code agent feature │ │ ├── ultra_loop/ # Autonomous agent loop │ │ └── ultra_think/ # 7-mode reasoning engine │ ├── cortex/ # Skill recommendation engine │ ├── orchestration/ # DAG-based workflow engine │ ├── multitenancy/ # Multi-tenant isolation │ └── llm/ # Multi-provider cognitive routing ├── jebat-npm/ # npx/npm wrapper package ├── db/ # SQLite RAG wiki & memory files └── docs/ # Platform architecture references ``` --- ## Safety Classification Tiers To safeguard target servers, JEBAT categorizes all tool execution into permission tiers: - **AUTO**: Read-only, safe commands executed without intervention (e.g. `cat`, `grep`). - **CONFIRM**: Write and modification commands prompting user validation (e.g. `write`, `patch`, `git commit`). - **DANGEROUS**: Destructive or privilege-escalating commands requiring explicit terminal validation tags (e.g. `rm -rf`, `sudo`). --- ## Platform Support | OS | Shell | Status | |----|-------|--------| | macOS | zsh/fish/bash | Full | | Linux | bash/zsh/fish | Full | | Windows | PowerShell/CMD/Git Bash | Full | | WSL2 | bash | Full | --- ## Documentation - [MCP integration](MCP.md) — Local IDE setup and self-hosted transport - [Workstation MCP guide](WORKSTATION_MCP_GUIDE.md) — IDE integration guide - [llama.cpp VPS deployment](docs/LLAMACPP_VPS_DEPLOYMENT.md) — Inference runbook - [Production Compose](infra/vps/vps/docker-compose.prod.yml) — Public API and WebUI stack - [Production Nginx vhost](infra/vps/vps/nginx.jebat.online.conf) — Public routing --- ## Model Fine-Tuning (JEBAT-Builder) Train a custom model that knows YOUR code style and JEBAT architecture. ### Free Cloud GPU Options | Platform | GPU | VRAM | Free Tier | Best For | |----------|-----|------|-----------|----------| | **Google Colab** | T4 | 15GB | ✅ Unlimited sessions | Quick experiments | | **Kaggle** | T4 x2 | 30GB | ✅ 30h/week | Medium fine-tunes | | **AWS SageMaker** | T4 | 16GB | ✅ 4h/session | Quick tests | | **Paperspace** | M4000 | 8GB | ✅ Limited hours | Development | | **Lightning AI** | T4 | 16GB | ✅ 22h/month | Development | | **GCP Free Tier** | T4/V100/A100 | Varies | ✅ $300 credit | Serious training | ### Paid Options (Cheapest) | Platform | GPU | Cost/Hour | Best For | |----------|-----|-----------|----------| | **RunPod** | RTX 3090 24GB | $0.45 | Budget training | | **Vast.ai** | RTX 3090 | $0.20-0.40 | Cheapest | | **Lambda** | A100 80GB | $1.10 | Production | | **Together.ai** | A100 80GB | $1.36 | Fine-tuning API | ### Fine-Tuning Steps ```bash # Step 1: Create training dataset from your repos python -m jebat.ml.create_dataset --repos /path/to/your/code # Step 2: Fine-tune on Google Colab (free T4 GPU) # - Use Unsloth for efficient LoRA (4-bit quantization) # - Base model: qwen2.5-coder:7b # - Time: 2-4 hours # - Cost: $0 # Step 3: Export as GGUF python -m jebat.ml.export_gguf --model ./finetuned_model --format Q4_K_M # Step 4: Deploy to jebat.online scp ./jebat-builder.Q4_K_M.gguf root@72.62.255.206:/tmp/ ssh root@72.62.255.206 "ollama create jebat-builder -f /tmp/Modelfile" ``` ### Your Hardware | Device | GPU | VRAM | Training Capability | |--------|-----|------|---------------------| | **Laptop** | RTX 3060 | 6GB | QLoRA for 3B-4B models | | **Server .206** | None (CPU) | N/A | Inference only | | **Google Colab** | T4 | 15GB | LoRA for 7B-13B models | | **Kaggle** | T4 x2 | 30GB | QLoRA for 7B-13B models | --- ## License & Organization Developed and maintained under strict data residency governance by NusaByte. Licensed under the MIT Open Source License. Built by Shaidan Shaari (humm1ngb1rd). 🗡️ **JEBAT** — *Because warriors remember everything that matters.*