colaworldmodel

GitHub Pages static reports for:

• the seed=42 Recurrent Transformer × TBPTT experiment; and
• the seed=42 COLA × LSTM World Models VizDoom TakeCover run.

Public URL: https://botwwt.github.io/colaworldmodel/

The site preserves the original RT/TBPTT report at index.html and adds the World Models report at world-models.html. It mirrors lightweight, public-facing experiment artifacts and intentionally excludes model checkpoints and large canonical datasets.

Local preview

python -m http.server 8080 --directory D:\DreamerV3\colaworldmodel

Then open http://127.0.0.1:8080/.

Reproducibility scope

• MNIST and CIFAR-10
• seed=42, n=1
• 15 epochs per cell
• K in {1, 4, 8, 16, 32}
• fixed 64-step shared RTM
• 10 completed formal cells

See artifacts/REPORT.md for the full Chinese report.

World Models result

• task: VizDoom TakeCover
• data: 10,000 episodes
• world model: 512-unit LSTM, 5-component MDN, COLA block_state, T_wm=500
• optimizer: Adam 1e-3, update-count-matched
• world-model training: 20 epochs, 1,140 steps, 57M transitions
• controller: CMA-ES, population 64, 16 dream rollouts/candidate, 200 generations
• final actual evaluation: 100 episodes, mean 184.57, sample SD 54.96

The learned controller scores 1975.28 on selected held-out dream validation but only 184.57 in the actual environment. This is recorded as a negative transfer result, not a successful reproduction. The report and downloadable artifacts are in world-models.html and artifacts/world_models/vizdoom_lstm_cola_seed42/.