MineChoice completed-house reward checkpoint 140
This model has not beaten Minecraft. No held-out full-game success rate or server-verified Ender Dragon kill exists for this checkpoint.
This immutable checkpoint is a 73-input, 45-action categorical MLP trained on real local Minecraft Java 1.20.4 reward transitions. It uses direct structured state, no pixels, and no teacher fallback. Coded game skills execute selected actions; the weights alone are not a standalone bot.
The natural seed 384455844 began with empty inventory and no
privileged setup in 20260928-061257-train-384455844. In continued-world episode 20260928-100156-train-persistent-002, the bot
finished one 52-part house at {'x': 185, 'y': 118, 'z': -111}. Its recorded reward transition
shows completed houses 0 to 1. The same episode crafted a replacement iron
pickaxe and stopped at that tool checkpoint. Its 54
real reward transitions trained update 140 (7084 cumulative
optimized transitions). This was not an uninterrupted fresh-world win or a
measured house-completion rate.
model.safetensors and policy.json contain identical neural weights.
config.json gives the exact feature and action order. milestone_evidence.json
records the lineage and milestone scope. Bundled policy files are the episode's
source snapshot. Training ran locally on an RTX 3090; inference can run locally
without a hosted service. No world save, credential, or raw gameplay log is included.
Recovery of the earlier lost diamonds, Nether entry, End entry, and game completion remain unverified for this checkpoint.