Download experiments/task-transfer/demon-attack/README.md from MLL-Lab/LAGEN-datasets: direct link, hf CLI and curl.
- Browser
- Download file 1.23 kB
-
https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/resolve/main/experiments/task-transfer/demon-attack/README.md
- Command line
-
hf download hf://datasets/MLL-Lab/LAGEN-datasets/experiments/task-transfer/demon-attack/README.md
-
curl -L -o README.md https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/resolve/main/experiments/task-transfer/demon-attack/README.md
Demon Attack task transfer
Two formal policies train on Demon Attack: zero latency at step 7000 and mixed latency at step 5000. Both evaluate on Demon Attack, Asterix, Atlantis and AirRaid at latency labels 0, 2 and 4, with 20 episodes per condition.
manifest.json pins checkpoints and original evaluation files. paper_comparison.json contains all 24 recomputed means and the paper values.
The corrected mixed evaluation is demon_attack_mixed_024_horizonfix_full_20261001_002008. The paper records mixed Demon Attack L0 as 5812; its public evaluation gives 5832.25. Other 23 points agree within integer display rounding. The original plotted CSV remains unchanged.
The paper names delays as raw frames. The corrected mixed configuration uses 60 environment FPS and 15 observation FPS: label 4 is 266.67 ms, or 16 raw frames. Original units remain in each source file. Exact checkpoint identity binding to these evaluations remains unverified.
The mixed checkpoint lacks its original training configuration, normalization statistics and exact training-data binding. Zero-policy evaluation summaries lack resolved configs and action traces. Missing assets remain explicit.