Qwen25 TOK-CC/Dist and Tokinv Sweep

Note

This is an observational follow-up page for the completed Qwen2.5 TOK-CC/dist-weight and token-inverse-surprisal sweeps.

Sweep Coverage

Qwen2.5 0.5B z8 collection: 6/6 completed.
Qwen2.5 0.5B z64 tokcc/dist collection: 3/3 completed.
Qwen2.5 ratio/tokinv follow-up collection: expected cells are tracked in results and data/followup_collection_status.csv.
Shared paired z8/z64 settings: dist_norm, tokcc_a0p25, dist_plus_tokcc_a0p25_w0p05_tokcc0p95, dist_plus_tokcc_a0p25_w0p5_tokcc0p5, dist_plus_tokcc_a0p25_w0p75_tokcc0p25, and tokinv_a0p25, all at b0p01/lin100.
Integrated z64-only additions include dist0p25_tokcc0p75_a0p25 and tokinv_a1_eps1em8 for Qwen2.5 0.5B/1.5B, plus 1.5B 50:50 and 75:25 ratio rows.

loss_setting	z8_matched_router_score	z64_matched_router_score	delta_z64_minus_z8_matched_router_score	z8_final_answer_accuracy	z64_final_answer_accuracy	z8_within_task_router_probe	z64_within_task_router_probe	z8_posterior_probe	z64_posterior_probe
dist_norm_b0p01_lin100	0.7141	0.7575	0.0434	0.9965	0.998	0.7171	0.76	0.6656	0.9642
tokcc_a0p25_b0p01_lin100	0.6950	0.7952	0.1003	0.9776	0.9938	0.7177	0.8092	0.6838	0.974
dist_plus_tokcc_a0p25_w0p05_tokcc0p95_b0p01_lin100	0.7182	0.577	-0.1411	0.9936	0.9991	0.7295	0.578	0.6831	0.7896
dist0p25_tokcc0p75_a0p25_b0p01_lin100		0.7388			0.9967		0.7447		0.9062
dist_plus_tokcc_a0p25_w0p5_tokcc0p5_b0p01_lin100	0.6467	0.6949	0.0482	0.9983	0.9983	0.6518	0.6965	0.6864	0.9351
dist_plus_tokcc_a0p25_w0p75_tokcc0p25_b0p01_lin100	0.7967	0.8502	0.0535	0.9969	0.995	0.8014	0.8574	0.8131	0.9913
tokinv_a0p25_b0p01_lin100	0.7259	0.5685	-0.1574	0.9964	0.9995	0.7301	0.5688	0.8483	0.8551
tokinv_a1_eps1em8_b0p01_lin100		0.6776			0.9991		0.6783		0.6927

Paired table: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-paper-method-grid-multi-task-z64-qwen25_20260501/data/followup_z8_z64_paired.csv
Integrated summary: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-paper-method-grid-multi-task-z64-qwen25_20260501/data/followup_summary.csv
Model comparison snapshot: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-paper-method-grid-multi-task-z64-qwen25_20260501/data/model_comparison_snapshot.csv
Generator: path:experiments/synthetic_sequences/disentangled/analysis/run_qwen25_paper_method_grid_multi_task_z64_20260501.py