Major changes from this session:
Training:
- 0.6.0 training running: 9B on steel141 3090 Ti, 27B on rented H100 NVL
- 7,256 merged training examples (up from 3,183)
- New training data: failure modes (85), midloop messaging (27),
prompt injection defense (29), personality (32), gold from quarantine
bank (232), new tool examples (30), claude's own experience (10)
- All training data RCON-validated at 100% pass rate
- Bake-off: gemma3:27b 66%, qwen3.5:27b 61%, translategemma:27b 56%
Oracle Bot (Mind's Eye):
- Invisible spectator bot (mineflayer) streams world state via WebSocket
- HTML5 Canvas frontend at mind.mortdec.ai
- Real-time tool trace visualization with expandable entries
- Streaming model tokens during inference
- Gateway integration: fire-and-forget POST /trace on every tool call
Reinforcement Learning:
- Gymnasium environment wrapping mineflayer bot (minecraft_env.py)
- PPO training via Stable Baselines3 (10K param policy network)
- Behavioral cloning pretraining (97.5% accuracy on expert policy)
- Infinite training loop with auto-restart and checkpoint resume
- Bot learns combat, survival, navigation from raw experience
Bot Army:
- 8-soldier marching formation with autonomous combat
- Combat bots using mineflayer-pvp, pathfinder, armor-manager
- Multilingual prayer bots via translategemma:27b (18 languages)
- Frame-based AI architecture: LLM planner + reactive micro-scripts
Infrastructure:
- Fixed mattpc.sethpc.xyz billing gateway (API key + player list parser)
- Billing gateway now tracks all LAN traffic (LAN auto-auth)
- Gateway fallback for empty god-mode responses
- Updated mortdec.ai landing page
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bake-off (0.5.0 vs 0.4.0):
- Overall: 46.8% vs 45.2% (+1.6%), 0 errors vs 2
- Enchantments: +47% (20% → 67%)
- EssentialsX: +60% (0% → 60%)
- Effects: +25% (0% → 25%)
- Regressions: fill_build -67%, world -20%
Knowledge Lookup Tools (4 new):
- plugin.docs_lookup: WorldGuard, WorldEdit, CoreProtect, EssentialsX, LuckPerms docs
- minecraft.changelog_lookup: version history from Minecraft Wiki
- paper.docs_lookup: Paper server-specific documentation
- Wired into gateway model-driven tool loop and exploration self-play
Exploration Self-Play:
- General (vanilla MC) and plugins focus modes
- Wiki-grounded: model researches before acting, validates through RCON
- 2,243 exploration examples generated, 150 kept after quality filtering
Training Progress Chart:
- SVG chart showing training examples and inverse loss across versions
- Added to MODEL_CARD.md for Gitea display
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
GPU Scheduler (gpu.sethpc.xyz):
- Live dashboard with 4 GPUs, training monitor, loss sparklines
- Preset-based job scheduler with 3 triggers (time, finish_training, cost)
- Model selection per GPU, pipeline configuration
- Tool self-play and training pipeline types
- Behind Google OAuth, live-refresh without page reload
Tool Architecture (14 tools):
- 3 new tools: world.nearby_entities, memory.read, memory.write
- 7 script.* tools: write, validate, execute, read, list, delete, schedule
- ScriptManager: full mcfunction datapack CRUD with RCON validation
- Training data: 1,430 tool examples (up from 1,159)
Plugin Deployment (paper-ai-25567):
- WorldGuard 7.0.12, CoreProtect CE 23.1, EssentialsX 2.21.2, Vault 1.7.3
- Fresh greenfield world reset
- 104 RCON-validated plugin training examples
Event Dispatcher:
- Watches server log for deaths, joins, advancements, PvP kills
- Configurable trigger probability and cooldowns per event type
- Deployed to dev server, fires god_system prompts on events
- 21 event-response training examples
Training Infrastructure:
- train_lora.py: --save-steps 50, --resume from checkpoint
- run_training.sh: stops Ollama, activates conda, restarts after
- Passwordless sudo for ollama services on steel141
- Dev server added to MCSManager with autoStart
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Versioning scheme: semantic versioning (MAJOR.MINOR.PATCH)
- 0.x.0 = pre-release development
- 1.0.0 = first public/monetized release
Renamed everywhere: PLAN.md, training scripts, self-play, overnight script,
status printer, whitelist app, discord bot, all training data references.
Ollama models retagged: mortdecai-v4 → mortdecai:0.4.0
Server configs updated on all three servers.
Self-play restarted with new model name.
Entity targeting + radius-aware kill + distance scale training added.
Seed dataset: 2,503 + tool: 1,159 + self-play: 5,059 = 8,721 total examples
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause: self-play opened/closed a new TCP socket for every RCON command
(hundreds/minute). Paper's RCON listener creates a thread per connection,
overwhelming the server until it stopped.
Fix: PersistentRCON class maintains a single connection per server with
auto-reconnect. Thread-safe via lock. Connection pool keyed by host:port.
Applied to:
- mc_aigod_paper.py (prod paper-ai + dev)
- mc_aigod.py (shrink-world)
- self_play.py (training data generation)
- persistent_rcon.py (shared module)
Before: ~100+ RCON connections/minute → server crash
After: 3 persistent connections total → stable
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bumped from 20 rounds/tier to 50. Reduced sleep from 1s to 0.1s.
GPUs should run near 100% — Ollama queues requests internally.
mortdecai-sites container (CT 650) created on pve112.
Landing page live at mortdec.ai.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Each cycle runs all three tiers at the same time on different GPUs:
- Tier 1 (drills) on GPU A
- Tier 2 (self-critique) on GPU B
- Tier 3 (adversarial) on GPU C
GPU assignments rotate each cycle for even wear.
3x throughput vs sequential. RCON handles concurrent commands.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Round-robin load balancing across three Ollama instances:
- 141:11434 (RTX 3090 Ti 24GB)
- 141:11435 (RTX 2080 Ti 11GB) — new second instance
- 179:11434 (RTX 4000 16GB)
Each tier cycles to a different GPU. 3x throughput overnight.
Cycles: Tier 1 drills → Tier 2 self-critique → Tier 3 adversarial → repeat
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Self-play (training/scripts/self_play.py):
- Model generates edge-case prompts across 9 categories
- Attempts commands via RCON, self-corrects on errors
- Successful traces → standard training examples
- Error correction traces → multi-turn tool-calling examples
- Anti-collapse: focuses on categories model is weakest in
- Ready for v4 deployment, not yet active
Qwen3.5-9B base model bake-off (147/1542 cases):
- 70.1% OK (vs 34% Qwen3-8B base) — 2x improvement
- 29.9% MISS (mostly God/prayer — no persona training)
- 15.6% needed syntax fixes
- Avg 7.5s response (thinking tokens)
- Strong v4 candidate: better base + tool-calling architecture
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- All 644 examples tagged: 0=blocked(15), 1=refuse(33), 2=warn(24), 3=normal(498), 4=generous(74)
- Training output now includes risk_level field for decision transparency
- Model learns to classify risk before generating commands
- Validator can sanity-check: risk 0-1 should have empty commands
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Swarm bots (ingame/swarm_bots.js):
- 10 survival bots with generated names (SwiftWolf, DarkWolf, etc.)
- All bots wander, take damage, auto-respawn, pray when hurt
- Gemini + Dolphin(5%) + Multilingual(3%) prompt generation
- 20-60s interaction interval per bot
Distillation results:
- 222 sudo examples via Haiku ($0.28)
- 122 god examples via Haiku ($0.37) — with God Soul personality
- Total: 344 distilled, $0.65 spent of $5 budget
- RCON validation: 74.7% fully valid, 30 real errors out of ~1000 commands
validate_distilled.py:
- Executes distilled commands on live server via RCON
- Distinguishes real errors from benign (no player online)
- Tags each example with validation status
Dev server switched to Claude Haiku via Anthropic API:
- llm_provider: anthropic with $5 budget cap
- Auto-fallback to Ollama when budget exhausted
- Cost tracking with logging
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
God Soul (agent/prompts/god_soul.md):
- Adapted from Claude's soul framework for the Minecraft God character
- Defines identity, principals hierarchy, decision-making framework
- Spectrum of responses (generous→silence), risk awareness, multilingual divinity
- Honesty within character, intervention guidelines
- Deployed to both prod and dev servers
System prompts updated:
- God prompt loads soul document dynamically
- Intervention prompt references soul for personality guidance
- Both include multilingual instruction (match player's language)
Distillation pipeline (training/scripts/distill.py):
- Sends all training examples through Claude API
- Haiku for sudo ($0.25), Sonnet for god ($0.50)
- Budget-capped, cost-tracked, --dry-run supported
- Outputs distilled.jsonl with Claude-quality responses
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Ingested 128 new examples from bot-driven data collection.
Dropped: 86 duplicates, 19 language mismatches, 10 prompt leaks, 19 empty.
Changed default epochs from 3 to 1 (previous run overfit at loss 0.10).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Switch fp16 to bf16 (RTX 3090 Ti is Ampere, supports BF16 natively)
- Include system prompt in training conversations (mode-aware: sudo/god/god_system)
- Include message field only for god modes
- Add determine_mode() and get_system_prompt() helpers
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- training/scripts/train_lora.py: Unsloth QLoRA trainer for qwen3:8b
- training/scripts/train_lora.sh: Launch script for steel141 RTX 3090 Ti
- eval/bakeoff.py: Fixed token budget (400->1500) that caused qwen3
models to exhaust tokens on thinking, added --no-think flag
- agent/serve.py: Default model changed to gemma3n:e4b
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bake-off tested 7 models on 31 seed examples via GPU-accelerated Ollama
on node-197 RTX 4000. gemma3n:e4b leads for serving (80.6% cmd match,
100% safety, 5.9s). qwen3:8b recommended as fine-tuning base (Apache 2.0,
best syntax quality, strong ecosystem). Full research in MODEL_RESEARCH.md.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- IDEA.md: project scope (Minecraft ops AI assistant via qwen3-coder LoRA/SFT)
- PLAN.md: complete roadmap with prior art analysis, architecture, phased plan, dev server docs
- data/schema.json: training example JSON Schema with negative_output support
- data/processed/seed_dataset.jsonl: 31 validated examples from repair code, prayer logs, session history
- data/validate_dataset.py: schema validator with summary statistics
- ingame/: Mineflayer bot framework (test_connect, spawn_bots, aware_bots with full event logging)
- Directory structure for knowledge/, eval/, training/, agent/ (Phase 1.3+ work)
Seth
Mortdecai
Seth