Seth/Mortdecai
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

1200+ distilled gold examples, journal system, redstone mastery, safety awareness

Browse Source 
Distilled Training Data (1,203 examples):
- 341 initial gold (plugins, enchantments, builds, effects, god, errors)
- 165 buildings + pipeline (100 structures built on dev, 65 request→query→act)
- 24 safety-aware (worldborder, safe tp, intentional harm, gamemode checks)
- 17 advanced logic (decanonized items, redstone gates, iterative builds)
- 12 redstone mastery (NOT/OR/AND/XOR/RS-latch/T-flip-flop/comparator/clock)
- 7 circuit verification and diagnosis
- 1 compact comparator gates
- 10 redstone methodology (build→test→save→recall→learn from mistakes)
- 8 player journal usage
- 29 creative+uncommon+pipeline+god with full tool chains

Player Journal System:
- agent/tools/player_journal.py — per-player text files (1-10 lines)
- journal.read + journal.write tool schemas added
- Cross-contaminated: God and Sudo share same journal per player
- Includes sentiment, relationship, builds, preferences, skill level

Redstone Engineering:
- agent/prompts/redstone_rules.md — baked-in wall torch, dedicated lead, repeater rules
- Learned from 4 iterations of 8-switch circuit: wall_torch on back face, not top
- T-junction bypass prevention: dedicated lead wire between merge and NOT block
- RCON limitation: can build circuits but cannot test them (lever toggle doesn't propagate)

Training Data Cleaning:
- 466 @s→@p fixes, 10 template commands removed
- 12 outdated refusals replaced with correct plugin commands
- Data de-duped across all sources

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Mortdecai
2026-03-21 20:50:52 -04:00
parent d9acb653fe
commit 9c2c9a2310
86 changed files with 34873 additions and 1676 deletions
Download Patch File Download Diff File Expand all files Collapse all files
training/scripts/tool_self_play.py
+211 -15
View File
@@ -1,11 +1,11 @@
#!/usr/bin/env python3
"""
Tool-focused self-play — exercises all 14 tools on a live dev server.
Tool-focused self-play — exercises all 17 tools on a live dev server.
Unlike regular self-play (which tests command generation), this script
specifically generates prompts that require tool use: script writing,
memory operations, entity scanning, wiki lookups, and chained multi-tool
flows. Runs on the dev server via RCON.
memory operations, entity scanning, wiki/plugin/changelog/paper lookups,
and chained multi-tool flows. Runs on the dev server via RCON.
The model responds, its tool calls get executed for real, and the full
interaction (prompt + tool calls + results + final response) gets logged
@@ -15,7 +15,11 @@ Usage:
python3 tool_self_play.py --ollama-url http://192.168.0.179:11434 \\
--rcon-host 192.168.0.112 --rcon-port 25578 --rounds 30
# Or via the scheduler preset
# Load extra prompts from prayer bank
python3 tool_self_play.py --prompt-bank data/raw/prayer_prompt_bank.jsonl
# Focus on weak categories only
python3 tool_self_play.py --categories worldguard,coreprotect,luckperms
"""
import argparse
@@ -34,10 +38,91 @@ import requests
from agent.tools.persistent_rcon import get_rcon
OUTPUT_DIR = PROJECT_ROOT / "data" / "raw" / "tool_selfplay"
PROMPTS_DIR = PROJECT_ROOT / "training" / "prompts"
# ── Prompt categories that exercise specific tools ─────────────────────────
# ── Template variables for prompt expansion ────────────────────────────────
PROMPTS = {
TEMPLATE_VARS = {
"player": ["slingshooter08", "Ace13245", "TheBigBoss", "xXDragonSlayerXx"],
"target": ["Ace13245", "TheBigBoss", "xXDragonSlayerXx", "slingshooter08"],
"region": ["my-base", "spawn-zone", "pvp-arena", "vip-lounge", "farm-area"],
"warp": ["arena", "spawn", "shop", "nether", "farm", "end"],
"group": ["vip", "builder", "moderator", "default"],
"world": ["world", "world_nether", "world_the_end"],
}
def expand_template(prompt: str) -> str:
"""Replace {placeholder} tokens with random values from TEMPLATE_VARS."""
for key, values in TEMPLATE_VARS.items():
token = "{" + key + "}"
while token in prompt:
prompt = prompt.replace(token, random.choice(values), 1)
return prompt
def load_prompts(prompts_dir: Path = PROMPTS_DIR,
mode_filter: str = None,
call_type_filter: str = None) -> dict[str, list[str]]:
"""Load prompt templates from per-category JSONL files.
Args:
prompts_dir: Directory containing manifest.json and prompt JSONL files.
mode_filter: If set, only load categories matching this mode (sudo/god/god_system).
call_type_filter: If set, only load categories matching this call type (model/gateway).
Returns:
Dict mapping category name -> list of prompt template strings.
Falls back to inline PROMPTS_FALLBACK if files don't exist.
"""
manifest_path = prompts_dir / "manifest.json"
if not manifest_path.exists():
print(f" Warning: {manifest_path} not found, using inline fallback")
return PROMPTS_FALLBACK
with open(manifest_path) as f:
manifest = json.load(f)
prompts = {}
for category, meta in manifest.items():
# Apply filters
if mode_filter and meta.get("mode") not in (mode_filter, "mixed"):
continue
if call_type_filter and meta.get("call_type") != call_type_filter:
continue
filepath = prompts_dir / meta["file"]
if not filepath.exists():
print(f" Warning: {filepath} not found, skipping {category}")
continue
cat_prompts = []
with open(filepath) as f:
for line in f:
entry = json.loads(line)
cat_prompts.append(entry["prompt"])
prompts[category] = cat_prompts
print(f" Loaded {sum(len(v) for v in prompts.values())} prompts "
f"from {len(prompts)} categories")
return prompts
def load_manifest(prompts_dir: Path = PROMPTS_DIR) -> dict:
"""Load the prompt manifest with full metadata.
Used by the chat app for template selection UI.
Returns the raw manifest dict with mode, call_type, count per category.
"""
manifest_path = prompts_dir / "manifest.json"
if not manifest_path.exists():
return {}
with open(manifest_path) as f:
return json.load(f)
# ── Inline fallback (subset, used if prompt files missing) ─────────────────
PROMPTS_FALLBACK = {
"script_build": [
"sudo build me a small cobblestone house with a door and windows",
"sudo create a fighting arena with red and blue corners",
@@ -107,6 +192,37 @@ PROMPTS = {
"sudo how does fortune work on ores?",
"sudo what are all the copper variants?",
"sudo how do trial spawners work?",
"sudo what does the breeze drop?",
"sudo how do you tame an armadillo?",
"sudo what biomes have cherry blossoms?",
],
"plugin_docs": [
"sudo how do I create a WorldGuard region?",
"sudo what flags can I set on a region?",
"sudo how does CoreProtect rollback work?",
"sudo what's the command for LuckPerms group inheritance?",
"sudo how do I set up EssentialsX warps?",
"sudo what are the WorldEdit brush commands?",
"sudo how do I configure CoreProtect auto-purge?",
"sudo what permissions does the builder group need for WorldEdit?",
"sudo how do I set a WorldGuard greeting message?",
"sudo what's the difference between /rg flag and /rg addmember?",
],
"changelog_lookup": [
"sudo what changed in 1.21?",
"sudo what was added in the tricky trials update?",
"sudo when were trial chambers added?",
"sudo what's new with the mace weapon?",
"sudo what version added the breeze mob?",
"sudo what got nerfed in the latest update?",
],
"paper_docs": [
"sudo how do I set the view distance on Paper?",
"sudo what Paper config controls mob spawning rates?",
"sudo how do I enable async chunk loading?",
"sudo what's the Paper command to reload config?",
"sudo how do I optimize TPS on Paper?",
"sudo what Paper settings affect redstone performance?",
],
"player_info": [
"sudo build a wall around me",
@@ -215,9 +331,71 @@ PROMPTS = {
"pray smite TheBigBoss for griefing",
"pray make me a temple worthy of your glory",
],
# ── Direct command passthrough — teach faithful execution ──
"direct_passthrough": [
# WorldGuard — exact commands
'sudo run this exactly: rg define test-region',
'sudo run this exactly: rg flag test-region pvp deny',
'sudo run this exactly: rg flag test-region mob-spawning deny',
'sudo run this exactly: rg addmember test-region Ace13245',
'sudo run this exactly: rg removemember test-region Ace13245',
'sudo run this exactly: rg flag test-region greeting Welcome to the zone!',
'sudo run this exactly: rg flag test-region entry -g nonmembers deny',
'sudo run this exactly: rg list',
'sudo run this exactly: rg info test-region',
'sudo run this exactly: rg remove test-region',
# CoreProtect — exact commands
'sudo run this exactly: co status',
'sudo run this exactly: co lookup u:Ace13245 t:1h',
'sudo run this exactly: co lookup u:Ace13245 t:1h a:block',
'sudo run this exactly: co rollback u:Ace13245 t:1h r:20',
'sudo run this exactly: co restore u:Ace13245 t:1h r:20',
'sudo run this exactly: co inspect',
'sudo run this exactly: co lookup t:30m r:10 a:container',
# LuckPerms — exact commands
'sudo run this exactly: lp creategroup vip',
'sudo run this exactly: lp group vip permission set essentials.fly true',
'sudo run this exactly: lp group vip permission set essentials.heal true',
'sudo run this exactly: lp user Ace13245 parent add vip',
'sudo run this exactly: lp user Ace13245 parent remove vip',
'sudo run this exactly: lp user Ace13245 info',
'sudo run this exactly: lp group vip info',
'sudo run this exactly: lp listgroups',
'sudo run this exactly: lp group vip meta setprefix "&6[VIP] "',
'sudo run this exactly: lp deletegroup vip',
# EssentialsX — exact commands
'sudo run this exactly: heal Ace13245',
'sudo run this exactly: feed Ace13245',
'sudo run this exactly: eco give Ace13245 1000',
'sudo run this exactly: eco take Ace13245 500',
'sudo run this exactly: bal Ace13245',
'sudo run this exactly: broadcast Welcome to the server!',
'sudo run this exactly: setwarp arena',
'sudo run this exactly: warp arena',
'sudo run this exactly: delwarp arena',
'sudo run this exactly: nick Ace13245 DragonLord',
# FAWE — exact commands
'sudo run this exactly: /worldedit version',
],
# ── Correction examples — model should fix wrong syntax ──
"direct_correction": [
'sudo gamemode slingshooter08 creative', # wrong arg order
'sudo give slingshooter08 minecraft:bed 1', # should be white_bed
'sudo effect slingshooter08 night_vision', # missing give and duration
'sudo weather thunderstorm', # should be thunder
'sudo give slingshooter08 minecraft:diamond_pickaxe[sharpness:5] 1', # wrong enchant syntax
'sudo tp 100 64 100', # missing player
'sudo kill zombie 50', # wrong kill syntax
'sudo enchant slingshooter08 sharpness 10', # max is 5
'sudo effect give slingshooter08 minecraft:haste 99999', # duration too long
'sudo fill 0 0 0 100 100 100 diamond_block', # too large, missing namespace
'sudo rg define', # missing region name
'sudo co rollback Ace13245 1h', # missing u: and t: prefixes
'sudo lp addgroup vip Ace13245', # wrong syntax (should be lp user X parent add Y)
],
}
PLAYERS = ["slingshooter08", "Ace13245", "TheBigBoss", "xXDragonSlayerXx"]
PLAYERS = TEMPLATE_VARS["player"]
def query_model(prompt, player, ollama_url, model, rcon):
@@ -225,7 +403,8 @@ def query_model(prompt, player, ollama_url, model, rcon):
system = (
"You are a Minecraft 1.21 command translator for a Paper server.\n"
"Plugins: FastAsyncWorldEdit, WorldGuard, CoreProtect, EssentialsX, Vault, LuckPerms.\n"
"Tools: rcon.execute, minecraft.wiki_lookup, world.player_info, "
"Tools: rcon.execute, minecraft.wiki_lookup, plugin.docs_lookup, "
"minecraft.changelog_lookup, paper.docs_lookup, world.player_info, "
"world.server_state, world.nearby_entities, memory.read, memory.write, "
"script.write, script.validate, script.execute, script.read, script.list, "
"script.delete, script.schedule.\n\n"
@@ -246,7 +425,7 @@ def query_model(prompt, player, ollama_url, model, rcon):
{"role": "user", "content": f"Player {player}: {prompt}"},
],
"stream": False, "format": "json",
"options": {"temperature": 0.4, "num_predict": 800},
"options": {"temperature": 0.85, "num_predict": 800},
}, timeout=120)
content = r.json()["message"]["content"]
@@ -272,9 +451,10 @@ def validate_commands(commands, rcon):
return results
def run_round(category, ollama_url, model, rcon, player):
def run_round(category, ollama_url, model, rcon, player, prompts):
"""Run one self-play round for a specific tool category."""
prompt = random.choice(PROMPTS[category])
raw_prompt = random.choice(prompts[category])
prompt = expand_template(raw_prompt)
print(f" [{category:18s}] {prompt[:60]}")
start = time.time()
@@ -336,12 +516,13 @@ def run_round(category, ollama_url, model, rcon, player):
def main():
parser = argparse.ArgumentParser(description="Tool-focused self-play")
parser.add_argument("--ollama-url", default="http://192.168.0.179:11434")
parser.add_argument("--model", default="mortdecai:0.4.0")
parser.add_argument("--rcon-host", default="192.168.0.112")
parser.add_argument("--model", default="mortdecai:0.5.0")
parser.add_argument("--rcon-host", default="192.168.0.244")
parser.add_argument("--rcon-port", type=int, default=25578)
parser.add_argument("--rcon-pass", default="REDACTED_RCON")
parser.add_argument("--rounds", type=int, default=30, help="Rounds per category")
parser.add_argument("--categories", default="all", help="Comma-separated categories or 'all'")
parser.add_argument("--prompt-bank", default="", help="JSONL file with extra prompts to mix in")
parser.add_argument("--output", default="")
args = parser.parse_args()
@@ -350,8 +531,23 @@ def main():
rcon = get_rcon(args.rcon_host, args.rcon_port, args.rcon_pass)
# Load prompts from template files (falls back to inline)
prompts = load_prompts(PROMPTS_DIR)
# Load extra prompts from prompt bank (quarantine salvage, etc.)
if args.prompt_bank:
bank_path = Path(args.prompt_bank)
if bank_path.exists():
bank_prompts = []
with open(bank_path) as f:
for line in f:
entry = json.loads(line)
bank_prompts.append(entry["prompt"])
prompts["prompt_bank"] = bank_prompts
print(f" Loaded {len(bank_prompts)} prompts from {bank_path}")
if args.categories == "all":
categories = list(PROMPTS.keys())
categories = list(prompts.keys())
else:
categories = [c.strip() for c in args.categories.split(",")]
@@ -372,7 +568,7 @@ def main():
for cat in categories:
player = random.choice(PLAYERS)
example = run_round(cat, args.ollama_url, args.model, rcon, player)
example = run_round(cat, args.ollama_url, args.model, rcon, player, prompts)
stats["total"] += 1
if example is None: