synthetic-pretrained-disentangled-from-entangled-v4

Status: partial. 19/24 cells are completed and analyzable, 3 failed, and 2 are still running.

This collection is a focused follow-up to the v2 and v3 from-entangled loss sweeps. It tests the low-KL limit more aggressively by introducing a zero-beta anchor, a very small constant-beta setting, and a matched mild posterior-variance comparison at beta=0.01 with full warmup. The current partial snapshot already shows strong task-dependent differences between the zero-beta and low-but-nonzero regimes, but the two low-beta base_conversion cells are still running.

• design
• results

Artifacts

• Notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/15_apr17_ss_from_entangled_v4_analysis.ipynb
• Aggregate directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v4/
• Figures directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v4/figures/

Context

• Prior design: synthetic-pretrained-disentangled-from-entangled-v3 design
• Prior results: synthetic-pretrained-disentangled-from-entangled-v3 results
• Earlier sweep: synthetic-pretrained-disentangled-from-entangled-v2 design

Runs

• Slurmkit collection: synthetic-pretrained-disentangled-from-entangled-v4
• Planned design: 6 tasks x 4 loss settings x 1 seed = 24 jobs
• Current state: 19 completed, 3 failed, 2 running
• Date range: 2026-04-17 to present