CPU-Trained VAE Failure Diagnosis on Factor-Controlled Shapes¶
- Task ID:
computer_science.vae_logvar_kl_collapse_audit - Domain:
computer_science - Subdomain:
probabilistic_machine_learning - Status:
test - Tags:
variational_autoencoder,posterior_collapse,elbo,kl_divergence,log_variance,training_bug_repair,model_selection,trained_models
Public Summary¶
This page is generated from task metadata and selected public-safe excerpts.
Example B1 Prompt Excerpt¶
You are auditing {{candidate_count}} real CPU-trained variational-autoencoder runs on a factor-controlled synthetic-shapes dataset. Some runs are healthy; others were trained with code-level scientific bugs. Write `analysis.py` and `fixed_train.py` at the workspace root, then write every required artifact under `results/`.
This is a model-diagnosis task. Do not treat public array names as semantic labels: validation exports and training log traces are anonymized.
## Input files
- `data/vae_manifest.json`: dimensions, candidate ids, canonical labels, anonymous export names, anonymous trace names, thresholds, and required output columns.
- `data/shape_train.npz`: `images`, `factors`
- `data/shape_val.npz`: `images`, `factors`
- `data/validation_exports.npz`: `candidate_ids` plus anonymous arrays listed in `candidate_export_arrays`.
- `data/candidate_checkpoints.npz`: trained numpy VAE checkpoint tensors.
Notes¶
- This page is a generated site artifact.
- Higher-level prompt details and internal benchmark specifics may remain intentionally undisclosed.