Datasets:

GoodStartLabs
/

hex-lora-opus-magnum-instructions-only-results-full

variant stringclasses 14 values	model_id stringclasses 5 values	model_size stringclasses 4 values	is_lora bool 2 classes	lora_path stringclasses 9 values	base_path stringclasses 5 values	representation stringclasses 3 values	task stringclasses 57 values	task_bucket stringclasses 3 values	sample_id stringclasses 57 values	epoch_k int64 0 17	solved bool 2 classes	best_reward float64 0 0.78	num_agent_turns int64 0 15	input_tokens int64 0 618k	output_tokens int64 0 46.6k
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	0	false	0	15	110,502	500
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	0	false	0	15	108,867	415
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	0	false	0	15	117,814	911
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	0	false	0	15	109,825	1,016
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	0	false	0	15	110,508	890
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	0	false	0	15	140,708	1,945
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	1	false	0	15	113,067	1,068
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	1	false	0	15	118,832	1,281
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	1	false	0	15	131,838	1,590
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	2	false	0	15	118,488	1,926
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	3	false	0.05	15	116,331	2,163
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	1	false	0	15	136,597	1,911
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	2	false	0	15	123,890	2,642
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	1	false	0	15	135,193	3,606
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	3	false	0	15	140,058	3,302
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	1	false	0	15	129,485	3,392
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	2	false	0	15	132,121	3,481
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	2	false	0.05	15	136,212	4,008
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	3	false	0.05	15	132,006	3,696
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	4	false	0	15	128,136	3,688
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	3	false	0.05	15	119,404	3,433
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	2	false	0	15	122,834	1,555
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	4	false	0	15	158,708	5,238
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	3	false	0.05	15	134,651	3,241
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	2	false	0	15	158,815	5,990
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	3	false	0	15	152,850	5,399
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	4	false	0	15	131,699	2,753
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed01	eval_fixed	gen010__seed01	0	false	0.05	15	126,303	1,801
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	0	false	0.05	15	122,938	1,528
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed01	eval_fixed	gen010__seed01	1	false	0	15	119,638	2,491
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p013_hangover_cure	campaign	p013_hangover_cure	5	false	0.05	15	169,773	7,153
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	1	false	0.05	15	122,453	1,838
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	0	false	0	15	118,537	1,470
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	1	false	0	15	120,554	1,769
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	2	false	0	15	137,952	2,378
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	0	false	0	15	118,230	924
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	1	false	0	15	112,412	614
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p016_hair_product	campaign	p016_hair_product	5	false	0	15	175,819	5,481
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	4	false	0	15	140,234	6,769
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p009_face_powder	campaign	p009_face_powder	5	false	0.05	15	201,507	8,539
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p011_waterproof_sealant	campaign	p011_waterproof_sealant	5	false	0.05	15	142,803	8,830
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed05	eval_fixed	gen010__seed05	0	false	0	15	127,459	599
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	4	false	0.05	15	169,729	7,320
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	3	false	0	15	135,987	3,136
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed05	eval_fixed	gen010__seed05	1	true	0.711864	7	59,310	1,066
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed01	eval_fixed	gen010__seed01	2	false	0	15	138,653	4,164
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	4	false	0	15	132,526	3,971
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	2	false	0	15	126,952	3,007
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed01	eval_fixed	gen010__seed01	3	false	0.05	15	204,071	5,286
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	2	false	0.05	15	134,268	4,000
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	3	false	0	15	146,686	3,712
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed05	eval_fixed	gen010__seed05	2	false	0	15	127,762	1,043
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	3	false	0	15	133,173	2,566
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	0	false	0	15	114,891	541
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	1	false	0.05	15	107,772	960
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed05	eval_fixed	gen010__seed05	3	false	0	15	132,767	2,182
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	2	false	0	15	129,163	1,710
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed01	eval_fixed	gen010__seed01	4	false	0.05	15	147,544	6,469
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p010_refined_gold	campaign	p010_refined_gold	5	false	0	15	219,641	10,157
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	4	false	0	15	144,135	4,887
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	4	false	0.05	15	128,493	3,524
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed02	eval_fixed	gen010__seed02	5	false	0	15	142,041	5,114
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	0	false	0	15	128,191	1,110
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed09	eval_fixed	gen010__seed09	0	false	0	15	124,828	579
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed09	eval_fixed	gen010__seed09	1	false	0	15	139,975	1,265
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed04	eval_fixed	gen010__seed04	5	false	0.05	15	192,290	6,078
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed03	eval_fixed	gen010__seed03	5	false	0.05	15	147,881	6,032
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	4	false	0.05	15	204,651	11,399
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed08	eval_fixed	gen010__seed08	0	false	0	15	144,523	1,998
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	3	false	0.05	15	138,882	3,953
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed11	eval_fixed	gen010__seed11	0	true	0.682482	2	12,894	323
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed08	eval_fixed	gen010__seed08	1	false	0.05	15	161,164	2,511
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	4	false	0	15	128,669	3,578
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed06	eval_fixed	gen010__seed06	5	false	0.05	15	134,144	3,859
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed10	eval_fixed	gen010__seed10	0	false	0	15	127,452	1,112
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed05	eval_fixed	gen010__seed05	4	false	0	15	142,168	5,024
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	1	false	0	15	140,417	3,476
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed09	eval_fixed	gen010__seed09	2	false	0	15	138,547	1,646
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	2	false	0	15	140,273	3,460
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed10	eval_fixed	gen010__seed10	1	false	0	15	125,080	1,945
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed09	eval_fixed	gen010__seed09	3	false	0	15	151,673	2,451
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	3	false	0	15	162,165	3,938
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	campaign_instructions_only__p007_stabilized_water	campaign	p007_stabilized_water	5	false	0.05	15	203,644	13,291
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed10	eval_fixed	gen010__seed10	2	false	0	15	126,344	2,414
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed08	eval_fixed	gen010__seed08	2	false	0	15	162,756	3,515
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed09	eval_fixed	gen010__seed09	4	false	0.05	15	160,028	4,331
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed12	eval_fixed	gen010__seed12	0	false	0	15	109,619	448
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed12	eval_fixed	gen010__seed12	1	false	0	15	119,576	824
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed12	eval_fixed	gen010__seed12	2	false	0	15	123,580	784
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed11	eval_fixed	gen010__seed11	1	false	0	15	117,638	1,549
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed12	eval_fixed	gen010__seed12	3	false	0	15	121,452	1,493
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	4	false	0	15	153,561	5,569
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed11	eval_fixed	gen010__seed11	2	false	0	15	128,781	2,583
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed11	eval_fixed	gen010__seed11	3	false	0	15	117,427	2,152
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed13	eval_fixed	gen010__seed13	0	false	0	15	109,738	520
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed12	eval_fixed	gen010__seed12	4	false	0.05	15	125,014	2,046
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed13	eval_fixed	gen010__seed13	1	false	0	15	119,749	1,057
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed07	eval_fixed	gen010__seed07	5	false	0	15	173,006	7,531
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed08	eval_fixed	gen010__seed08	3	false	0.05	15	175,449	6,512
4b-base	Qwen/Qwen3.5-4B	4B	false	null	tinker-base://Qwen/Qwen3.5-4B	n/a	eval_fixed_instructions_only__gen010__seed10	eval_fixed	gen010__seed10	3	false	0.05	15	162,398	5,339

End of preview. Expand in Data Studio

hex-lora-opus-magnum-instructions-only-results-full

Extended-attempt sister of hex-lora-opus-magnum-instructions-only-results on the same 57-puzzle "instructions-only" eval set (Opus Magnum campaign + curated holdout puzzles), with k=18 attempts/task for 4B/27B/Gemini variants and k=12 for K2.6 variants (vs k=6 in the original release). Total: 12,312 trajectories across 14 variants.

The agent runs an interactive Python REPL and must submit() a working .solution file to the in-game verifier.

Files

samples.jsonl — flat per-trajectory scoring + metadata (one row per attempt across all variants). 12,312 rows.
14 .eval files — one per variant, each combines its original-release run plus all overnight *-extra and *-extra2 runs into a single inspect-ai log with epoch_k re-numbered 0..(k−1) across all attempts. Open with inspect view <file> or inspect_ai.log.read_eval_log(...).

Top-line accuracy (combined-k, partial-credit reward)

Sorted high → low by pass rate. n = total trajectories per variant.

variant	k	solved/n	pass	mean_r
gemini-flash	18	510/1026	49.7%	0.427
27b-lora-dsl	18	452/1026	44.1%	0.385
kimi-textarena-64tok-200i	12	300/684	43.9%	0.382
27b-lora-python	18	437/1026	42.6%	0.375
kimi-lora-dsl	12	245/684	35.8%	0.327
kimi-lora-python	12	239/684	34.9%	0.317
27b-base	18	355/1026	34.6%	0.329
kimi-base	12	229/684	33.5%	0.301
kimi-textarena-1024tok-50i	12	156/684	22.8%	0.220
gemini-flash-lite	18	212/1026	20.7%	0.210
4b-reasoning-2048tok-500i	12	57/684	8.3%	0.102
4b-lora-python	18	73/1026	7.1%	0.085
4b-lora-dsl	18	64/1026	6.2%	0.076
4b-base	18	41/1026	4.0%	0.051

pass = fraction of attempts that reached solved=True from the in-game verifier. mean_r is mean of best_reward per attempt under the partial-credit reward profile (cost-based score for solved attempts; ≤ 0.5 partial credit for unsolved-but-progressed attempts).

Why more k

Doubling-to-tripling attempts/task was specifically to tighten 95% CIs and enable paired-bootstrap LoRA-vs-base tests. Two top-line findings out of the extra data that were unresolved at k=6:

All four OM-trained dsl/python LoRAs significantly beat their base models (paired bootstrap, 95% CI excludes 0): 27B-dsl +9.4 pp, 27B-py +8.0 pp, 4B-py +3.1 pp, 4B-dsl +2.2 pp.
K2.6 textarena-64tok LoRA also beats K2.6 base by +10.4 pp paired — the largest LoRA-vs-base effect in the sweep — while textarena-1024tok significantly underperforms K2.6 base by 11 pp.

The unpaired cluster-bootstrap CI doesn't resolve any of these LoRA-vs-base comparisons even at k=18 — the paired structure is what produces the significance.

Per-variant attempt counts (combined run-dir provenance)

For traceability:

variant	run_dirs combined	k
4b-base	opus-4b-base-k6-mt4096 + extra + extra2	18
4b-lora-python	opus-4b-py-step300-k6-mt4096 + extra + extra2	18
4b-lora-dsl	opus-4b-dsl-step300-k6-mt4096 + extra + extra2	18
4b-reasoning-2048tok-500i	qwen3p5-4b-reasoning-2048tok-500i-k6 + extra	12
27b-base	opus-27b-base-tinker-k6-mt1536 + extra + extra2	18
27b-lora-python	opus-27b-py-step210-k6-mt4096 + extra + extra2	18
27b-lora-dsl	opus-27b-dsl-step210-k6-mt4096 + extra + extra2	18
kimi-base	kimi-k2.6-tinker-base-k6-mt1536 + extra	12
kimi-lora-python	opus-k26-py-step150-k6 + extra	12
kimi-lora-dsl	opus-k26-dsl-step150-k6 + extra	12
kimi-textarena-64tok-200i	kimi-k26-textarena-64tok-200i-k6 + extra	12
kimi-textarena-1024tok-50i	kimi-k26-textarena-1024tok-50i-k6 + extra	12
gemini-flash	gemini-flash-latest-or-k6-mt4096 + extra + extra2	18
gemini-flash-lite	gemini-3.1-flash-lite-or-k6-mt1536 + extra + extra2	18

extra/extra2 runs use identical agent config, dataset, scoring profile, and (where relevant) Tinker URI as their parent run. They are independent re-evaluations using fresh sampling; differences in pass rate within a variant across runs reflect sampling noise (cluster-CI ~10 pp at k=6).

Eval task

Same 57-puzzle opus-magnum-sean-modal-k1-sum set as the original release, split as: 6 campaign puzzles, 37 eval-fixed generated holdout puzzles, 14 "minimal" handcrafted micro-puzzles.

Agent uses the in-house REPL harness (agents/om_repl_runner.py) with:

param	value
`max_turns`	15
`max_tokens`	1536 (K2.6 LoRAs / 27B-base mt=1536 / Gemini Flash Lite) / 4096 (4B family / 27B LoRAs / Gemini Flash)
`temperature`	1.0 (Tinker) / OR provider default with thinking off
`n_attempts` (per run)	6
`scoring_profile`	`partial-credit`
`enable_thinking`	off (all variants)

Schema (`samples.jsonl`)

field	type	notes
`variant`	str	one of the 14 variant names above
`model_id`	str	HF id (`Qwen/Qwen3.5-4B`, `Qwen/Qwen3.5-27B`, `moonshotai/Kimi-K2.6`, `google/gemini-flash-latest`, `google/gemini-3.1-flash-lite-preview`)
`model_size`	str	`4B` / `27B` / `Kimi` / `Gemini`
`is_lora`	bool	True if LoRA-finetuned
`lora_path`	str\|null	Tinker sampler-checkpoint URI for LoRA variants
`base_path`	str\|null	Sampling endpoint for base/Gemini variants
`representation`	str	`python` / `dsl` / `n/a`
`task`	str	full task dir name, e.g. `campaign_instructions_only__p010_refined_gold`
`task_bucket`	str	`campaign` / `eval_fixed` / `minimal`
`sample_id`	str	task name with bucket prefix stripped, e.g. `p010_refined_gold`
`epoch_k`	int	attempt index in `0..k−1` (k=18 or k=12 depending on variant)
`solved`	bool	scorer verdict from in-game verifier
`best_reward`	float	partial-credit reward in `[0, 1]`
`num_agent_turns`	int	how many assistant messages the trial used
`input_tokens`, `output_tokens`	int	aggregated over the trial
`final_assistant_text`	str	empty here — full transcripts in `.eval` files

Loading

from datasets import load_dataset
ds = load_dataset("path/to/samples.jsonl", split="samples")
print(ds[0])

# Pass rate by variant
import polars as pl
df = pl.read_ndjson("samples.jsonl")
print(
    df.group_by(["variant"])
      .agg(pl.col("solved").mean().alias("pass"),
           pl.col("solved").sum().alias("n_solved"),
           pl.len().alias("n"))
      .sort("pass", descending=True)
)

To replay a single trajectory with the inspect viewer:

pip install inspect-ai
inspect view 27b-lora-python.eval

Provenance

Generated by scripts/run_local_repl_eval.py in opus-magnum-rle on 2026-05-03 / 2026-05-04 / 2026-05-05. Per-run summaries and trajectory JSONs live under jobs/local/<run-id>/ in that repo. Inspect-ai logs combined via scripts/build_k18_results_dir.py.

License

MIT. Underlying base model licenses (Qwen, Kimi, Google) apply to model outputs.

Downloads last month: 154