synthetic-pretrained-disentangled-from-entangled-v1

Status: complete - 5/6 runs analyzed; sorting_algorithms failed with CUDA OOM.

This collection tests whether initializing disentangling from the entangled pretrained source model yields cleaner latent strategy structure than direct-pretrained disentangling alone. The headline result is asymmetric: posterior-side structure improves clearly on some tasks, especially base_conversion and linear_equation_solving, while router-side gains remain partial.

• design
• results
• analysis

Artifacts

• notebook: path:experiments/synthetic_sequences/disentangled/analysis/notebooks/12_apr14_ss_from_entangled_v1_analysis.ipynb
• aggregates: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v1
• exploratory figures: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v1/figures

Context

• depends on ss-disentangled-direct-pretrained-v1 as the direct-pretrained baseline
• followed by synthetic-pretrained-disentangled-from-entangled-v2

Runs

• slurmkit collection: path:.slurmkit/collections/synthetic-pretrained-disentangled-from-entangled-v1.yaml
• run count: 5/6 complete, 1/6 failed
• date range: 2026-04-14