azalio/laya-doom-map03

Laya Doom v0.3.1: MAP01–MAP03

One fixed Laya bundle completed FreeDoom MAP01, MAP02 and MAP03 without deaths in one sequential real-time series. The model chooses the action, target, weapon and combat movement. The controller routes, aims and presses buttons to execute those choices.

Map	Seed / skill	Exit time	Deaths	Full-decision HTTP p50 / p90
MAP01 → MAP02	48 / 3	168.400 s	0	266.01 / 288.46 ms
MAP02 → MAP03	54 / 3	420.400 s	0	295.81 / 337.32 ms
MAP03 → MAP04	54 / 3	347.114 s	0	309.47 / 367.67 ms

Recorded on macOS/MPS, float32, 35 game ticks per second. Every recording includes 105 ticks of the next map. Video, real-time pacing and model authority checks passed. MAP02 and MAP03 also passed navigation thresholds. MAP01 retains a warning: 32.23 seconds without a new region exceeds the 25-second threshold, so its overall quality check remains false.

These maps and seeds were used during development. This is one final series after multiple failed experiments, not an independent generalization or reliability benchmark. Actual RTT is included; responses are applied as soon as received, without the older artificial minimum delay of 16 ticks. There were no external model API calls. Local compute and training cost were not estimated. This release has no new Jev comparison.

Videos and telemetry · Verification and hashes · Reproduction and training.

Architecture and training

The eight checkpoints share one identical frozen encoder at inference. Question names select the trained head. Command, item, combat movement and mechanism heads also use learned numeric residual networks over observed facts. Their weights and feature specifications are covered by the model manifest. This is an architecture extension and supervised adaptation of Laya, not the unchanged upstream model.

Training uses recorded game observations, labels from offline policies, synthetic examples, and retention of previous model decisions. This is supervised imitation learning / behavior cloning, with no reward-based RL. The offline policies are not called by the game or inference server. Training and validation are correlated same-map development data.

The final command repair updates existing numeric layers while retaining all 1096 recorded decisions from earlier successful MAP01/MAP03 runs. The selected epoch is 65 of 1600; validation labels match in 1179/1191 cases. These numbers are training diagnostics, not gameplay success rates.

Bundle manifest: 74eea1043aa584bf825fe25ac40be5fc7affc555c63dea8712b9b807ae9cafb6. Use question-heads.json and SHA256SUMS to verify the files. training-assets/ contains the parent checkpoints, semantic caches and retention traces for the documented final training stages. It is optional for inference. Training datasets and scripts are in the GitHub release tag.

Base: convaiinnovations/laya/typed-decisions at 1c5edc17a7acd8701df6fc341c0d179f1c62c982; Laya source commit 42626c348753fbb17572a813127df2278a1ec527. Apache-2.0; see NOTICE.

Download

Install the repository and dependencies using the linked reproduction guide. From its root directory:

hf download azalio/laya-doom-map03 --revision v0.3.1 \
  --exclude 'training-assets/*' --local-dir checkpoints/laya-doom-v031
(cd checkpoints/laya-doom-v031 && shasum -a 256 -c SHA256SUMS)

The guide includes the exact server flags, three-map regression command, and training commands with the published dependencies. The v0.3.0 tag preserves the previous release, whose MAP01/MAP02 regression failed.