update cache

Signed-off-by: Yotam-Perlitz <[email protected]>

cache/aggregate_scoress_cache_2916d5df57069c497d825c9f1fac0bd4.csv

@@ -0,0 +1,62 @@
+model,score
+claude_3_5_sonnet_20240620,1.0
+gpt_4o_2024_05_13,0.9833333333333333
+gpt_4_0125_preview,0.9666666666666667
+gpt_4o_2024_08_06,0.95
+athene_70b,0.9333333333333333
+gpt_4o_mini,0.9166666666666666
+gemini_1_5_pro_api_preview,0.9
+mistral_large_2407,0.8833333333333333
+llama3_1_405b_instruct,0.8666666666666667
+glm_4_0520,0.85
+yi_large,0.8333333333333334
+deepseek_coder_v2,0.8166666666666667
+claude_3_opus_20240229,0.8
+gemma_2_27b_it,0.7833333333333333
+llama3_1_70b_instruct,0.75
+glm_4_0116,0.75
+glm_4_air,0.7333333333333333
+gpt_4_0314,0.7166666666666667
+gemini_1_5_flash_api_preview,0.7
+qwen2_72b_instruct,0.6833333333333333
+claude_3_sonnet_20240229,0.6666666666666666
+llama3_70b_instruct,0.65
+claude_3_haiku_20240307,0.6333333333333333
+gpt_4_0613,0.6166666666666667
+mistral_large_2402,0.6
+mixtral_8x22b_instruct_v0_1,0.5833333333333334
+qwen1_5_72b_chat,0.5666666666666667
+phi_3_medium_4k_instruct,0.55
+command_r_plus,0.5333333333333333
+mistral_medium,0.5166666666666667
+internlm2_5_20b_chat,0.5
+phi_3_small_8k_instruct,0.48333333333333334
+mistral_next,0.4666666666666667
+gpt_3_5_turbo_0613,0.45
+dbrx_instructruct_preview,0.43333333333333335
+internlm2_20b_chat,0.4166666666666667
+claude_2_0,0.4
+mixtral_8x7b_instruct_v0_1,0.38333333333333336
+gpt_3_5_turbo_0125,0.36666666666666664
+yi_34b_chat,0.35
+starling_lm_7b_beta,0.3333333333333333
+claude_2_1,0.31666666666666665
+llama3_1_8b_instruct,0.3
+snorkel_mistral_pairrm_dpo,0.2833333333333333
+llama3_8b_instruct,0.26666666666666666
+gpt_3_5_turbo_1106,0.25
+gpt_3_5_turbo_0301,0.23333333333333334
+gemini_1_0_pro,0.21666666666666667
+snowflake_arctic_instruct,0.2
+command_r,0.18333333333333332
+phi_3_mini_128k_instruct,0.16666666666666666
+tulu_2_dpo_70b,0.15
+starling_lm_7b_alpha,0.13333333333333333
+mistral_7b_instruct,0.11666666666666667
+gemma_1_1_7b_it,0.1
+llama_2_70b_chat,0.08333333333333333
+vicuna_33b_v1_3,0.06666666666666667
+gemma_7b_it,0.05
+llama_2_7b_chat,0.03333333333333333
+gemma_1_1_2b_it,0.016666666666666666
+gemma_2b_it,0.0

cache/aggregate_scoress_cache_e798cd9f99d4c09af5b81734eecc6592.csv

@@ -1,13 +1,122 @@
 model,score
-gpt_4o_2024_05_13,1.0
-qwen2_72b_instruct,0.8437710437710438
-gemma_2_27b_it,0.8225108225108225
-llama3_70b_instruct,0.7288840788840788
-qwen1.5_110b_chat,0.6958874458874459
-command_r_plus,0.5340909090909091
-qwen1.5_72b_chat,0.510909090909091
-llama3_8b_instruct,0.2897065897065897
-mixtral_8x7b_instruct_v0.1,0.27759740259740256
-command_r,0.1515151515151515
-qwen1.5_7b_chat,0.06897546897546898
-llama_2_7b_chat,0.0
+gpt_4o_2024_05_13,0.9847612958226769
+claude_3_5_sonnet_20240620,0.982905982905983
+gpt_4o_2024_08_06,0.9575873827791986
+gpt_4_turbo_2024_04_09,0.9428463693169576
+gpt_4_0125_preview,0.9171132221004344
+mistral_large_2407,0.8868286445012787
+llama3_1_405b_instruct,0.8672150411280846
+yi_large_preview,0.8641553641553642
+hermes_3_llama3_1_70b,0.8626160990712074
+smaug_qwen2_72b_instruct,0.8593911248710011
+claude_3_opus_20240229,0.8573567665639277
+llama3_1_70b_instruct,0.8528408270971201
+athene_70b,0.8493788819875776
+deepseek_coder_v2,0.8444160272804775
+qwen2_72b_instruct,0.8354710666091739
+yi_large,0.8346273291925466
+gpt_4_0613,0.8146763722211293
+llama3_70b_instruct,0.8127546753337573
+llama3_70b,0.8105600539811066
+gemma_2_27b_it,0.8045273029120115
+gpt_4o_mini_2024_07_18,0.8032033326150972
+gemma_2_9b_it_dpo,0.790057915057915
+llama3_instruct_8b_simpo,0.7884068278805121
+phi_3_5_moe_instruct,0.7808307533539731
+qwen1_5_110b_chat,0.776004448721167
+qwen1_5_32b,0.7658569500674763
+yi_1_5_34b_chat,0.7553884711779449
+llama_2_70b,0.7303193882141251
+mixtral_8x22b_instruct_v0_1,0.7256023690940907
+gemma_2_9b_it_simpo,0.7199248120300753
+qwen1_5_32b_chat,0.7149122807017544
+mixtral_8x22b_v0_1,0.7135490753911806
+yi_34b,0.7128879892037787
+internlm2_5_20b_chat,0.6842105263157895
+phi_3_small_128k_instruct,0.66937564499484
+phi_3_medium_4k_instruct,0.6675079642841117
+claude_3_sonnet_20240229,0.653911731916847
+gemma_2_9b_it,0.6422797189051059
+infinity_instruct_3m_0625_llama3_8b,0.6273115220483642
+mistral_v0_1_7b,0.6239316239316239
+phi_3_5_mini_instruct,0.6202270381836945
+mistral_medium,0.6122209165687427
+mistral_large_2402,0.6058211467418628
+claude_instant_1_2,0.6049896049896051
+claude_2_0,0.6020066889632107
+yi_1_5_9b_chat,0.5881787802840435
+qwen1_5_14b,0.5770917678812416
+command_r_plus,0.5761033510394125
+llama_65b,0.5736992052781527
+gpt_3_5_turbo_0613,0.5724018332713985
+qwen1_5_72b_chat,0.5668371367348349
+phi_3_mini_4k_instruct,0.5548245614035088
+deepseek_llm_67b_chat,0.5506756756756757
+claude_3_haiku_20240307,0.549424005945745
+yi_34b_chat,0.5455449728905107
+dbrx_instructruct,0.5344129554655871
+jurassic_2_jumbo_178b,0.532051282051282
+llama3_1_8b_instruct,0.5175232440678665
+claude_2_1,0.5110980545763154
+qwen2_7b_instruct,0.5034227726178191
+mistral_small_2402,0.49924585218702866
+mixtral_8x7b_v0_1,0.49324324324324326
+glm_4_9b_chat,0.46499582289055974
+qwen1_5_14b_chat,0.4621068436857911
+phi_3_small_8k_instruct,0.45481670929241264
+gpt_3_5_turbo_0301,0.4528985507246377
+snorkel_mistral_pairrm_dpo,0.4521151586368978
+gemma_7b,0.4471997300944669
+gpt_3_5_turbo_0125,0.4401920188365201
+llama3_8b,0.43302968960863697
+dbrx_instruct,0.4266409266409266
+llama3_8b_instruct,0.420135922511747
+phi_3_mini_128k_instruct,0.4153205904787544
+llama_2_13b,0.41490478332583597
+jurassic_2_grande_17b,0.39529914529914534
+openhermes_2_5_mistral_7b,0.3832617447168531
+mistral_7b_v0_3,0.3737553342816501
+mixtral_8x7b_instruct_v0_1,0.3713078251895724
+qwen1_5_7b,0.3508771929824561
+yi_1_5_6b_chat,0.3354636591478697
+falcon_40b,0.32812265707002547
+command_r,0.32386140074759
+internlm2_chat_20b,0.32252252252252256
+mistral_7b_v0_2,0.31970128022759603
+luminous_supreme_70b,0.30128205128205127
+starling_lm_7b_alpha,0.29823530624445954
+yi_6b,0.29234143049932526
+mistral_7b_instruct_v0_2,0.28609513981031004
+zephyr_7b_alpha,0.2838442157327606
+zephyr_7b_beta,0.2666234345800909
+gemma_1_1_7b_it,0.26226051061156724
+mistral_7b_instruct_v0_3,0.2537839697282422
+starling_lm_7b_beta,0.25234441602728047
+llama_2_7b,0.2391288049182786
+luminous_extended_30b,0.2329059829059829
+alpaca_7b,0.22072072072072071
+vicuna_33b_v1_3,0.2056404230317274
+phi_2,0.20087901666849037
+qwen2_1_5b_instruct,0.19711042311661506
+yi_6b_chat,0.1938854489164087
+qwen1_5_7b_chat,0.1916569245052217
+tulu_2_dpo_70b,0.17624223602484473
+qwen1_5_4b_chat,0.1674406604747162
+llama_2_70b_chat,0.15527950310559005
+gpt_neox_20b,0.14400584795321636
+vicuna_7b_v1_5,0.13619501854795973
+falcon_40b_instruct,0.13264580369843526
+gemma_7b_it,0.12136319058515854
+falcon_7b,0.11407257459889038
+gpt_j_6b,0.10160818713450293
+luminous_base_13b,0.08333333333333333
+llama_2_7b_chat,0.08304448781801049
+gemma_1_1_2b_it,0.07665903890160183
+olmo_7b,0.06545209176788123
+gemma_2b_it,0.05921052631578947
+qwen1_5_1_8b_chat,0.059167526659786716
+qwen2_0_5b_instruct,0.059081527347781215
+pythia_12b,0.054093567251461985
+pythia_6_9b,0.019736842105263157
+falcon_7b_instruct,0.013513513513513514
+qwen1_5_0_5b_chat,0.013157894736842105