Qwen25 Multi-Task Dynamics: Qwen2.5 0.5B

Note

This is an observational dynamics deep-dive for qwen25_05b. This page uses final-test metrics for ranking because final metrics are available.

Selection

Rows are ranked by within-task matched router score, computed as within-task router probe times strategy-match rate. The selected row is dist_plus_tokcc_a0p25_w0p75_tokcc0p25_b0p01_lin100 with score 0.8502 from final_test.

selected	collection_label	continuous_latent_dim	loss_setting	analysis_status	metric_source	matched_router_score	final_answer_accuracy	within_task_router_probe	all_task_router_probe	strategy_match_rate	posterior_probe	reconstruction_loss	kl_prior_loss	latest_global_step
selected	0.5B tokcc/dist sweep z64	64	dist_plus_tokcc_a0p25_w0p75_tokcc0p25_b0p01_lin100	complete_final	final_test	0.8502	0.995	0.8574	0.2242	0.9916	0.9913	0.0006	2.3351	1563
	0.5B tokcc/dist sweep z8	8	dist_plus_tokcc_a0p25_w0p75_tokcc0p25_b0p01_lin100	complete_final	final_test	0.7967	0.9969	0.8014	0.2016	0.9941	0.8131	0.0018	2.1244	1563
	0.5B paper grid z64	64	tokcc_a0p25_b0p01_lin100	complete_final	final_test	0.7952	0.9938	0.8092	0.2704	0.9827	0.974	0.5019	5.0621	1563
	0.5B paper grid z64	64	dist_norm_b0p01_lin100	complete_final	final_test	0.7575	0.998	0.76	0.2028	0.9968	0.9642	0.0108	1.5638	1563
	ratio/tokinv follow-up z64	64	dist0p25_tokcc0p75_a0p25_b0p01_lin100	complete_final	final_test	0.7388	0.9967	0.7447	0.2035	0.9921	0.9062	0.0147	2.5455	1563
	0.5B tokcc/dist sweep z8	8	tokinv_a0p25_b0p01_lin100	complete_final	final_test	0.7259	0.9964	0.7301	0.1963	0.9942	0.8483	0.5487	2.6247	1563
	0.5B tokcc/dist sweep z8	8	dist_plus_tokcc_a0p25_w0p05_tokcc0p95_b0p01_lin100	complete_final	final_test	0.7182	0.9936	0.7295	0.1854	0.9844	0.6831	0.0008	1.7038	1563
	raw beta baseline z64	64	raw_b0p01_const	complete_final	final_test	0.7175	0.9992	0.7181	0.1659	0.9992	0.7839	0.0321	1.9304	1563
	0.5B tokcc/dist sweep z8	8	dist_norm_b0p01_lin100	complete_final	final_test	0.7141	0.9965	0.7171	0.183	0.9958	0.6656	0.0027	1.7813	1563
	raw beta baseline z64	64	raw_b0p1_const	complete_final	final_test	0.7026	0.9989	0.7034	0.1678	0.9989	0.769	0.0323	5.836	1563
	0.5B tokcc/dist sweep z8	8	tokcc_a0p25_b0p01_lin100	complete_final	final_test	0.695	0.9776	0.7177	0.1941	0.9684	0.6838	0.53	1.9834	1563
	0.5B tokcc/dist sweep z64	64	dist_plus_tokcc_a0p25_w0p5_tokcc0p5_b0p01_lin100	complete_final	final_test	0.6949	0.9983	0.6965	0.2004	0.9976	0.9351	0.0172	1.6825	1563
	raw beta baseline z64	64	raw_b1_const	complete_final	final_test	0.684	0.9971	0.6867	0.149	0.9961	0.776	0.0331	1.7719	1563
	ratio/tokinv follow-up z64	64	tokinv_a1_eps1em8_b0p01_lin100	complete_final	final_test	0.6776	0.9991	0.6783	0.1407	0.9989	0.6927	0.5819	78.6067	1563
	raw beta baseline z64	64	raw_b1_lin100	complete_final	final_test	0.6708	0.9984	0.672	0.1603	0.9982	0.7899	0.0327	3.342	1563
	0.5B tokcc/dist sweep z8	8	dist_plus_tokcc_a0p25_w0p5_tokcc0p5_b0p01_lin100	complete_final	final_test	0.6467	0.9983	0.6518	0.1821	0.9921	0.6864	0.0075	1.4014	1563
	0.5B paper grid z64	64	tokcc_a0p25_b0p1_lin100	complete_final	final_test	0.592	0.9996	0.5924	0.1509	0.9994	0.7764	0.6479	2.6485	1563
	0.5B paper grid z64	64	dist_plus_tokcc_a0p25_w0p05_tokcc0p95_b0p01_lin100	complete_final	final_test	0.577	0.9991	0.578	0.148	0.9984	0.7896	0.0321	0.7372	1563
	0.5B paper grid z64	64	dist_norm_b0p1_lin100	complete_final	final_test	0.573	0.9995	0.5733	0.1261	0.9995	0.7207	0.032	0.0086	1563
	0.5B paper grid z64	64	raw_b0p01_lin100	complete_final	final_test	0.5719	0.9995	0.5723	0.1366	0.9994	0.7768	0.032	0.6675	1563
	0.5B paper grid z64	64	dist_plus_tokcc_a0p25_w0p05_tokcc0p95_b0p1_lin100	complete_final	final_test	0.5686	0.9994	0.5692	0.1208	0.999	0.7266	0.0321	0.0103	1563
	0.5B tokcc/dist sweep z64	64	tokinv_a0p25_b0p01_lin100	complete_final	final_test	0.5685	0.9995	0.5688	0.1251	0.9995	0.8551	0.5858	0.0191	1563
	0.5B paper grid z64	64	raw_b0p1_lin100	complete_final	final_test	0.565	0.9995	0.5653	0.1235	0.9995	0.6859	0.032	0.0062	1563

Validation Dynamics

The validation dynamics panel plots final-answer accuracy, matched router score, within-task router probe, all-task router probe, strategy-match rate, posterior-mu probe, reconstruction loss, and KL prior by epoch progress for each available run directory.

Router-Sampled Z-Space

Task-separated view: GIF columns are effective tasks; frame rows are effective tasks.

Task-separated Router-sampled z-space PCA animation for Qwen2.5 0.5B

Task-separated Router-sampled z-space PCA animation frames for Qwen2.5 0.5B

Source frame stream: path:.runs/synthetic_sequences/disentangled/synthetic-pretrained-disentangled-from-entangled-v2-tokcc-dist-weight-sweep-multi-task-z64-qwen25-05b/task-multi_task_src-ent-qwen25-05b_adapt-lora_lt-continuous_loss-recon-dist-0p75-tokcc-0p25_beta-0p01-_hcc4ede439fe42ed8_s314159/continuous_val_router_sampled_z_frames.jsonl

Posterior-Mu Z-Space

Selected Posterior-mu z-space PCA animation frames for Qwen2.5 0.5B

Task-separated view: GIF columns are effective tasks; frame rows are effective tasks.

Task-separated Posterior-mu z-space PCA animation for Qwen2.5 0.5B

Smoothed task-separated Posterior-mu z-space PCA animation for Qwen2.5 0.5B

Task-separated Posterior-mu z-space PCA animation frames for Qwen2.5 0.5B

Run-Level Notes

The selected run directory is path:.runs/synthetic_sequences/disentangled/synthetic-pretrained-disentangled-from-entangled-v2-tokcc-dist-weight-sweep-multi-task-z64-qwen25-05b/task-multi_task_src-ent-qwen25-05b_adapt-lora_lt-continuous_loss-recon-dist-0p75-tokcc-0p25_beta-0p01-_hcc4ede439fe42ed8_s314159. The data backing this page are in path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-paper-method-grid-multi-task-z64-qwen25_20260501/data/dynamics_qwen25_05b_metrics.csv and path:experiments/synthetic_sequences/disentangled/analysis/aggregates/synthetic-pretrained-disentangled-from-entangled-v2-paper-method-grid-multi-task-z64-qwen25_20260501/data/dynamics_qwen25_05b_eval_history.csv. The generation script is path:experiments/synthetic_sequences/disentangled/analysis/run_qwen25_paper_method_grid_multi_task_z64_20260501.py.

Qwen25 Multi-Task Dynamics: Qwen2.5 0.5B

Selection

Validation Dynamics

Router-Sampled Z-Space

Posterior-Mu Z-Space

Run-Level Notes

Links to this page: