Experiment Sources
Experiment Sources
Back to hub · single-task results · multi-task results
This note documents the concrete experiment collections and comparison aggregates that feed the paper-selected reconstruction-stack report. The directory name still reflects the original Apr26 paper-selection handoff, but the current report is a paper-grid synthesis over later Apr27, Apr30, May01, and May02 random-init collections, with earlier Apr24/Apr26 and qwen25 0.5B reports used as design context.
Integrated Paper-Grid Aggregate
The main single-task results note and the multitask note are generated from path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429. The current aggregate has 504 complete final-test cells: 432 single-task cells and 72 multi_task cells. It has no incomplete or failed final-test cells. Analogical summaries exist for all 432 single-task cells and all 72 multi_task cells.
| source collection | role in report | added settings | final-test cells | analogical status | notes |
|---|---|---|---|---|---|
exp2_continuous_cvae-Apr27-paper-basic-grid | Base paper-grid collection plus Apr30 appended artifacts in the aggregate inventory | raw, dist_norm, example_inv_surp, tokcc_a0p25, dist_plus_tokcc_a0p25, tokinv_a0p25, dist0p05_tokcc0p95_a0p25; each over six tasks plus multi_task, three beta settings, and two beta schedules | 294 | Complete for all 294 cells | The Apr30 slurmkit extension is listed separately below because it is a separate submitted collection, but its completed artifacts are folded into the Apr27 aggregate source inventory rather than a separate artifact_collection_name bucket. |
exp2_continuous_cvae-Apr30-paper-basic-grid-extension | Extension that completed missing paper-grid method families | tokinv_a0p25 and dist0p05_tokcc0p95_a0p25 over the same beta/schedule/task layout | included in Apr27 bucket | Complete where represented in current aggregate | Slurmkit source path: path:.slurmkit/collections/exp2_continuous_cvae-Apr30-paper-basic-grid-extension.yaml. |
exp2_continuous_cvae-May01-tokeninv-stress-dist-tokcc-followup | Follow-up stress test and intermediate dist:TOK-CC ratio expansion | tokinv_a1_eps1em8, tokinv_a1_eps1em4, dist0p10_tokcc0p90_a0p25, dist0p25_tokcc0p75_a0p25 over the same beta/schedule/task layout | 168 | Complete for all 168 cells | This collection supplies the 0.25 distribution-normalized / 0.75 TOK-CC comparison and the alpha-1 token-inverse-surprisal stress settings. |
exp2_continuous_cvae-May02-dist0p75-tokcc0p25-followup | Focused ratio follow-up | dist0p75_tokcc0p25_a0p25 over the same beta/schedule/task layout | 42 | Complete for all 42 cells | This collection supplies the 0.75 distribution-normalized / 0.25 TOK-CC comparison. Its final-test and analogical metrics are included in results and multitask. |
Repo-Side Artifacts
| artifact | path | used by |
|---|---|---|
| Main aggregate directory | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429 | results, dynamics, dynamics_tokinv_a0p25, tokcc_deepdive |
| Main aggregate report | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/report.md | Source for curated single-task result tables and figures |
| Main aggregate brief | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/brief.yaml | Machine-readable scope for the integrated paper-grid aggregate |
| Main aggregate findings | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/findings.json | Machine-readable headline findings |
| Multi-task side report | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/multitask | multitask |
Contextual Comparison Sources
These sources informed design and analysis, but they are not part of the 504-cell Apr27+Apr30+May01+May02 aggregate used by the current single-task and multi-task result tables.
| source | role | path |
|---|---|---|
| Apr24+Apr26 random-init unified reconstruction-stack aggregate | Historical random-init evidence used to choose the preliminary paper-facing loss subset and to track low-beta raw controls | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426 |
| qwen25 0.5B from-entangled reconstruction-stack aggregate | Pretrained/from-entangled comparison evidence used in the original design discussion | path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_from_entangled_v2_v4_reconstruction_stack_qwen25_05b_20260426 |
| qwen25 1.5B from-entangled collection | Planned or separate scale-up comparison; not integrated into this report's current random-init paper-grid aggregate | path:.slurmkit/collections/synthetic_sequences/families/disentangled_from_entangled/continuous_from_entangled_v2_v4_reconstruction_stack_qwen25_15b.yaml |
Naming Note
The directory name paper-selected-reconstruction-stack-losses_20260426 now describes the starting point rather than the current scope. A more accurate future rename would identify this as the random-init paper-grid reconstruction-stack synthesis, while keeping the date as lineage metadata if desired.
