WoW 3.3.5a Server — AzerothCore + Humanlike Playerbots

A private WotLK (3.3.5a, build 12340) server whose world is populated by 500 autonomous player bots: they quest, grind, travel, rest at inns, queue for battlegrounds and dungeons, upgrade their gear — and behave like people:

• 75 chat personalities with per-personality talkativeness, verbosity, and chaos, speaking through a locally fine-tuned LLM (wow-chat) in an authentic 2008 voice.
• Playstyles: a bot's personality drives its gameplay — grinders farm mobs, socializers loiter in town, idlers nap at inns, pvpers hunt world PvP.
• Gear awareness: bots inspect players they talk to and react in-voice — helpful types offer real spare items from their bags, merchants try to sell, elitists mock, and everyone recognizes raid sets and PvP resilience gear.
• Relationships: per-pair sentiment evolves through chat tone, duels, grouping, and guild life. Bots in good standing occasionally offer to group with you — most likely when they can confirm you share their quest.
• Arena coordination: bot arena teams focus one kill target (yours, if you attack first) and hold burst cooldowns until a teammate — including you — pops theirs. (Activates once bots level to 70+.)
• World PvP respect: fighting well beside bots earns their respect, saving one earns a lot (and a grateful "ty for the save"), and ganking a lowbie in front of them costs it. Bots judge each other's conduct too. Rising respect feeds the same guild/help/gift gates — so honorable PvP wins you friends.

New here? Start with docs/GETTING-STARTED.md — overview + install guide. Deep dive on every behavior system: docs/BOT-BEHAVIOR.md. Function-level code reference for every subsystem: docs/internals/. Full build/migration history with every problem and fix: docs/BUILD-NOTES.md.

Current host (since 2026-07-02): i7-14700K, 32 GB RAM, RX 7900 XT — the realm and its LLM inference run on the same box (Ollama/ROCm, ~114 tok/s, ~0.2 s per chat reply). Originally built on an i7-1185G7 box and migrated; the realm address is 127.0.0.1.

Stack

Component	What / where
Core	liyunfan1223/azerothcore-wotlk, branch Playerbot (custom fork — required, stock AzerothCore will not compile the bot module)
Bots	liyunfan1223/mod-playerbots, branch master + our patches
LLM chat	DustinHendrickson/mod-ollama-chat + our patches + local Ollama (ROCm) serving wow-chat (fine-tuned Qwen3-4B, Q8_0)
Database	MariaDB 11.8 (Debian packages) — needs the local patches
Client data	wowgaming/client-data release v19 (maps/vmaps/mmaps/dbc)

Directory layout

/home/admin/git/wow/
├── README.md                     # this file (operating reference)
├── docs/BOT-BEHAVIOR.md          # every bot system, in depth
├── docs/BUILD-NOTES.md           # chronological journal, problems + fixes
├── DEPENDENCIES.txt              # every apt package + external dependency
├── build.sh                      # clone (first run) + compile + install
├── start.sh                      # tmux-mode start/stop/status (AC> consoles)
├── reset-world.sh                # one-shot world wipe + reseed (see below)
├── restart-world.sh              # crash-restart loop (legacy; systemd covers this)
├── personalities.sql             # 42 custom personalities + playstyle mapping
├── patches/                      # ALL changes to the nested repos (see its README)
├── finetune/                     # wow-chat dataset generator + eval harness
├── gpu-box/                      # Ollama serving + QLoRA training scripts
├── systemd/                      # wow-auth / wow-world units (installed + enabled)
├── azerothcore-wotlk/            # sources (fork + modules/), build/ inside
└── server/                       # runtime install (CMAKE_INSTALL_PREFIX)
    ├── bin/                      # authserver, worldserver, *.log
    ├── etc/                      # worldserver.conf, authserver.conf
    │   └── modules/              # playerbots.conf, mod_ollama_chat.conf
    └── data/                     # Cameras/ dbc/ maps/ mmaps/ vmaps/

Running the server

Two ways — don't mix them (each refuses to start over the other):

systemd (default: boot autostart, crash restart, no console):

sudo systemctl start|stop|restart|status wow-auth wow-world
journalctl -u wow-world -f

Manual with consoles (when you want the AC> console):

sudo systemctl stop wow-auth wow-world   # if units are running
./start.sh            # both servers in tmux
./start.sh status     # processes / ports / sessions
./start.sh stop       # graceful shutdown
tmux attach -t world  # console (detach: Ctrl-b d)

MariaDB and Ollama are separate systemd services (also on boot).

Start the world over: ./reset-world.sh — asks for confirmation, then one-shot: wipes auth/characters/playerbots DBs, restarts, loads all 75 personalities before the first bot rolls, recreates the GM account, sets the realm row (name + current IP + loopback localAddress), and relaunches authserver if it loses the first-boot DB race. --no-start wipes and leaves the realm down; --full also re-imports the static world DB. Bot re-seeding takes minutes on this box (the docs' "10-40 min" was the old machine).

Secrets / user-specific settings live in .env (gitignored — copy .env.example): GM account name + password, realm name. Scripts fall back to the example defaults (admin / changeme123 / Gigi) when no .env exists.

Playing

• This box: launch "World of Warcraft (Gigi)" from the app menu / Desktop shortcut. The shortcut runs launch-client.sh, which self-heals before starting the game: this box is on DHCP, so the IP can drift ("unable to connect" at login) — the launcher rewrites the client's realmlist.wtf to the current IP, fixes acore_auth.realmlist + bounces authserver if stale, and clears the client's Cache/WDB. (reset-world.sh also derives the IP dynamically now.) Re-create the shortcut any time — e.g. after moving the repo — with ./install-shortcut.sh. Client is configured for 4K borderless (gxMaximize) with max UI scale; Config.wtf carries the OpenGL + windowed fixes that prevent the post-ToS freeze under Wine.
• Accounts: admin (GM 3, initial password changeme123 — change it: console account set password admin <new> <new>; defaults configurable via .env). Create more: account create <name> <pass> in the console or .account create in-game.
• Realm addressing: the local client connects to gigi.local — pinned to 127.0.0.1 in /etc/hosts — so it is immune to DHCP address changes; the realm row's localAddress=127.0.0.1 routes loopback clients back to loopback for the world server. Other LAN devices use the box's current IP (the launcher keeps the realm row's external address up to date). Auth on 3724, world on 8085, both bind 0.0.0.0. For WAN play, port-forward both and point acore_auth.realmlist.address at your public IP/DDNS.

Databases

MariaDB 11.8, user acore / password acore (localhost only): acore_auth, acore_world, acore_characters, acore_playerbots. Schema updates apply at worldserver startup (Updates.AutoSetup=1). InnoDB tuning: /etc/mysql/mariadb.conf.d/60-azerothcore.cnf (6G buffer pool).

Handy queries (Debian root uses unix_socket — sudo mariadb):

SELECT COUNT(*) FROM acore_characters.characters WHERE online=1;   -- bots in world
SELECT level, COUNT(*) FROM acore_characters.characters WHERE online=1 GROUP BY level;
SELECT playstyle, COUNT(*) FROM acore_characters.mod_ollama_chat_personality_templates GROUP BY playstyle;

Configuration highlights

Live values are the .conf files next to their .conf.dist templates; diff against the dist to see every local change. Full explanation of the behavior keys: docs/BOT-BEHAVIOR.md.

server/etc/modules/playerbots.conf:

• MinRandomBots / MaxRandomBots = 500, fresh-realm leveling (DisableRandomLevels = 1, RandombotStartingLevel = 1 — lowercase b is real), EnableNewRpgStrategy = 1 (the living-world brain), BG/LFG/auto-gear on.
• Playstyle profiles (ours): AiPlayerbot.RpgStatusProbWeight.<Profile>.<Status> overrides; defaults documented in the conf.
• Quest-help invites (ours): AiPlayerbot.QuestHelp{SentimentThreshold,ConfirmedChance,NearbyChance,RandomChance}.
• Rated arena (upstream, enabled): RandomBotAutoJoinBGRatedArena2v2Count = 2, 3v3Count = 1; teams auto-create when 70+ captains exist.

server/etc/modules/mod_ollama_chat.conf:

• Model = wow-chat, Url = http://127.0.0.1:11434/api/generate, MaxConcurrentQueries = 8, EnableTypingSimulation = 1, OllamaChat.RequestTimeout = 120 (our conf key).
• ChatPromptTemplate includes {gear_context} (our placeholder — bots inspect the player they talk to).
• EnableRPPersonalities = 1; sentiment tracking + our event nudges (SentimentEvent{Duel,Group,Guild}Adjustment).
• Bots only chatter near real players, so Ollama is idle until someone logs in.

⚠️ Local patches to the nested repos

The nested clones (core fork + both modules) carry substantial local changes, captured as three patch files in patches/ with apply/regenerate instructions in patches/README.md:

1. azerothcore-mariadb-compat.patch — build/run on MariaDB instead of MySQL 8.
2. mod-ollama-chat-enhancements.patch — MariaDB fix, per-personality behavior columns, sentiment event nudges, {gear_context}, playstyle migration, guild-hook registration bugfix, RequestTimeout.
3. mod-playerbots-playstyles.patch — playstyle profiles, arena coordination, quest-help invites.

After any git pull in the nested repos, re-apply before rebuilding. Rebuild: cd azerothcore-wotlk/build && make -j8 install (incremental; full toolchain in DEPENDENCIES.txt), then restart worldserver.

Debugging tip: a worldserver "segfault at 0 … error 6" in dmesg is AzerothCore's WPFatal assertion, not memory corruption — the real message is in server/bin/Errors.log.

The wow-chat model & fine-tuning

wow-chat (Ollama alias → Q8_0 quant of our QLoRA fine-tune of Qwen3-4B-Instruct) is what the bots speak through. The previous model is always kept as wow-chat:v1 — rollback is ollama cp wow-chat:v1 wow-chat, no server restart needed.

Retraining (new personalities, new context types): edit finetune/generate_dataset.py, regenerate, train on this box (~55 min), and follow the staged deploy in finetune/README.md — one training round produced a corrupted model with a perfectly clean loss curve; the coherence gate exists for a reason.

Monitoring & tuning

• Console server info → Update time diff: <50 ms great (this box idles at 1 ms with 500 bots).
• tail -f server/bin/Playerbots.log — bot activity; grep "\[Playstyle\]", "\[QuestHelp\]", "\[Arena\]" for our systems; Errors.log — problems.
• journalctl -u ollama -f — chat generations when players are near bots.
• Scale: raise MinRandomBots/MaxRandomBots (headroom for 1000+ here); new bot characters are created automatically on next boot.

Rough resource picture

500 bots: 1–5 ms update diff; worldserver RAM grows toward ~8–11 GB as bots explore (maps never unload). Ollama holds wow-chat resident (~5 GB VRAM, OLLAMA_KEEP_ALIVE=-1); QLoRA training peaks ~17.5 GB VRAM — evict resident models first (ollama stop wow-chat). MariaDB buffer pool capped at 6 GB.