Merge arena hard leaderboard and leaderboard table csv file

README.md

@@ -1,5 +1,5 @@
 ---
-title: Chatbot Arena Leaderboard
+title: LMSys Chatbot Arena Leaderboard
 emoji: 🏆🤖
 colorFrom: indigo
 colorTo: green
@@ -7,8 +7,7 @@ sdk: gradio
 pinned: false
 license: apache-2.0
 tags:
-- leaderboard
-sdk_version: 4.44.1
+  - leaderboard
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -6,7 +6,6 @@ import glob
 import re
 import gradio as gr
 
-
 def load_demo(url_params, request: gr.Request):
     logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")
     return basic_component_values + leader_component_values
@@ -42,7 +41,7 @@ def build_demo(elo_results_file, leaderboard_table_file):
         css=block_css,
     ) as demo:
         leader_components = build_leaderboard_tab(
-            elo_results_file, leaderboard_table_file, arena_hard_file, show_plot=True, mirror=True
+            elo_results_file, leaderboard_table_file, show_plot=True, mirror=True
         )
     return demo
 
@@ -63,10 +62,6 @@ if __name__ == "__main__":
     leaderboard_table_files = glob.glob("leaderboard_table_*.csv")
     leaderboard_table_files.sort(key=lambda x: int(x[18:-4]))
     leaderboard_table_file = leaderboard_table_files[-1]
-    
-    arena_hard_files = glob.glob("arena_hard_auto_leaderboard_*.csv")
-    arena_hard_files.sort(key=lambda x: float(x[29:32]))
-    arena_hard_file = arena_hard_files[-1]
 
     demo = build_demo(elo_result_file, leaderboard_table_file)
     demo.launch(share=args.share, server_name=args.host, server_port=args.port)

arena_hard_auto_leaderboard_v0.1.csv

@@ -1,61 +0,0 @@
-model,score,rating_q025,rating_q975,CI,avg_tokens,date
-gpt-4-turbo-2024-04-09,82.63,80.75,84.6,"(1.9, 2.0)",662.0,2024-07-31
-claude-3-5-sonnet-20240620,79.35,77.25,80.62,"(2.1, 1.3)",567.0,2024-07-31
-gpt-4o-2024-05-13,79.21,77.42,80.71,"(1.8, 1.5)",696.0,2024-07-31
-gpt-4-0125-preview,77.96,75.94,79.9,"(2.0, 1.9)",619.0,2024-07-31
-athene-70b-0725,76.83,74.84,78.74,"(2.0, 1.9)",683.0,2024-07-31
-gpt-4o-mini-2024-07-18,74.94,72.66,77.07,"(2.3, 2.1)",668.0,2024-07-31
-gemini-1.5-pro-api-0514,71.96,69.62,74.62,"(2.3, 2.7)",676.0,2024-07-31
-yi-large-preview,71.48,69.02,73.37,"(2.5, 1.9)",720.0,2024-07-31
-mistral-large-2407,70.42,68.11,72.43,"(2.3, 2.0)",623.0,2024-07-31
-llama-3.1-405b-instruct,64.09,61.43,66.55,"(2.7, 2.5)",633.0,2024-07-31
-glm-4-0520,63.84,61.28,66.19,"(2.6, 2.3)",636.0,2024-07-31
-yi-large,63.7,61.76,65.86,"(1.9, 2.2)",626.0,2024-07-31
-deepseek-coder-v2,62.3,59.82,64.72,"(2.5, 2.4)",578.0,2024-07-31
-claude-3-opus-20240229,60.36,57.56,62.34,"(2.8, 2.0)",541.0,2024-07-31
-gemma-2-27b-it,57.51,55.11,60.12,"(2.4, 2.6)",577.0,2024-07-31
-llama-3.1-70b-instruct,55.73,52.85,58.2,"(2.9, 2.5)",628.0,2024-07-31
-glm-4-0116,55.72,53.83,58.16,"(1.9, 2.4)",622.0,2024-07-31
-gemini-1.5-pro-api-0409-preview,53.37,51.13,56.66,"(2.2, 3.3)",478.0,2024-07-31
-glm-4-air,50.88,48.62,53.21,"(2.3, 2.3)",619.0,2024-07-31
-gpt-4-0314,50.0,50.0,50.0,"(0.0, 0.0)",423.0,2024-07-31
-gemini-1.5-flash-api-0514,49.61,47.46,52.17,"(2.1, 2.6)",642.0,2024-07-31
-qwen2-72b-instruct,46.86,44.57,49.29,"(2.3, 2.4)",515.0,2024-07-31
-claude-3-sonnet-20240229,46.8,44.12,49.04,"(2.7, 2.2)",552.0,2024-07-31
-llama-3-70b-instruct,46.57,43.84,49.18,"(2.7, 2.6)",591.0,2024-07-31
-claude-3-haiku-20240307,41.47,39.57,44.02,"(1.9, 2.6)",505.0,2024-07-31
-gpt-4-0613,37.9,35.6,40.36,"(2.3, 2.5)",354.0,2024-07-31
-mistral-large-2402,37.71,34.81,39.77,"(2.9, 2.1)",400.0,2024-07-31
-mixtral-8x22b-instruct-v0.1,36.36,34.21,38.55,"(2.1, 2.2)",430.0,2024-07-31
-qwen1.5-72b-chat,36.12,33.88,38.15,"(2.2, 2.0)",474.0,2024-07-31
-phi-3-medium-4k-instruct,33.37,31.26,35.14,"(2.1, 1.8)",517.0,2024-07-31
-command-r-plus,33.07,30.85,35.12,"(2.2, 2.0)",541.0,2024-07-31
-mistral-medium,31.9,29.66,34.31,"(2.2, 2.4)",485.0,2024-07-31
-phi-3-small-8k-instruct,29.77,27.94,31.97,"(1.8, 2.2)",568.0,2024-07-31
-mistral-next,27.37,25.4,29.09,"(2.0, 1.7)",297.0,2024-07-31
-gpt-3.5-turbo-0613,24.82,22.54,26.29,"(2.3, 1.5)",401.0,2024-07-31
-dbrx-instruct-preview,24.63,22.33,26.83,"(2.3, 2.2)",415.0,2024-07-31
-claude-2.0,23.99,21.71,25.65,"(2.3, 1.7)",295.0,2024-07-31
-mixtral-8x7b-instruct-v0.1,23.4,21.38,25.41,"(2.0, 2.0)",457.0,2024-07-31
-gpt-3.5-turbo-0125,23.34,21.67,25.27,"(1.7, 1.9)",329.0,2024-07-31
-yi-34b-chat,23.15,20.75,24.7,"(2.4, 1.6)",611.0,2024-07-31
-starling-lm-7b-beta,23.01,20.81,24.66,"(2.2, 1.6)",530.0,2024-07-31
-claude-2.1,22.77,20.65,25.43,"(2.1, 2.7)",290.0,2024-07-31
-llama-3.1-8b-instruct,21.34,19.71,23.09,"(1.6, 1.8)",861.0,2024-07-31
-snorkel-mistral-pairrm-dpo,20.73,19.04,22.05,"(1.7, 1.3)",564.0,2024-07-31
-llama-3-8b-instruct,20.56,18.82,22.61,"(1.7, 2.1)",585.0,2024-07-31
-gpt-3.5-turbo-1106,18.87,17.06,20.58,"(1.8, 1.7)",285.0,2024-07-31
-gpt-3.5-turbo-0314,18.05,16.57,20.06,"(1.5, 2.0)",334.0,2024-07-31
-gemini-pro,17.8,15.96,19.32,"(1.8, 1.5)",322.0,2024-07-31
-snowflake-arctic-instruct,17.61,16.12,19.27,"(1.5, 1.7)",365.0,2024-07-31
-command-r,17.02,15.73,18.51,"(1.3, 1.5)",432.0,2024-07-31
-phi-3-mini-128k-instruct,15.43,13.94,17.02,"(1.5, 1.6)",609.0,2024-07-31
-tulu-2-dpo-70b,14.99,13.05,16.82,"(1.9, 1.8)",550.0,2024-07-31
-starling-lm-7b-alpha,12.8,11.23,14.5,"(1.6, 1.7)",483.0,2024-07-31
-mistral-7b-instruct,12.57,11.05,14.11,"(1.5, 1.5)",541.0,2024-07-31
-gemma-1.1-7b-it,12.09,10.61,13.43,"(1.5, 1.3)",341.0,2024-07-31
-llama-2-70b-chat,11.55,10.02,13.01,"(1.5, 1.5)",595.0,2024-07-31
-vicuna-33b,8.63,7.59,9.84,"(1.0, 1.2)",451.0,2024-07-31
-gemma-7b-it,7.47,6.5,8.6,"(1.0, 1.1)",378.0,2024-07-31
-gemma-1.1-2b-it,3.37,2.74,4.14,"(0.6, 0.8)",316.0,2024-07-31
-gemma-2b-it,3.0,2.33,3.67,"(0.7, 0.7)",369.0,2024-07-31

elo_results_20240722.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c01b5a786ffd8bf64b72d39ca3fc7dee5483852c6a515f515532c8096a6e16d4
-size 3054518

elo_results_20240725.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1bfde449e424f12c9316f0c64062fc1ef9926e4a924bbaeafa455fcde0decb6f
-size 3073542

elo_results_20240730.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8db59f5f0852f26600f85c3188547e5d97b4641906c979bcc178cd5be7a7554c
-size 3137995

elo_results_20240731.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9aaff49a611fef8ff7bec058827d7de3426973c15bc538ac6bb39e764cc14b34
-size 3157482

elo_results_20240801.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:cdaa16f762adc54cbe05ab03dff88587491193852846b749980294b7f1ea2bec
-size 3182414

elo_results_20240805.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e638bfe870e4409c4d768d57287cc8f6310caf85191d8020ba89f7f59ee9f6d8
-size 3202114

elo_results_20240806.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a4f31f36636a280589bb1039f9d1f405a989df6a8f74d1af30555d891b23a416
-size 3261205

elo_results_20240813.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b4c5dfd4247b704f07e61ae27dc0642a3d3cfa9a6872cc3dc03d1888a594de9f
-size 2943734

elo_results_20240822.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c733894de1252232a63ce3632abe52504fb6bcf43e17bb49fea2b5ad8d76116f
-size 3004697

elo_results_20240823.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ec103aa6bf5d0f02f8bd2c69c8ccfc8f1be1b44c7dc004d967c8d5ce470975b5
-size 3039588

elo_results_20240827.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:fbcf63bc492b9e2018fdd2c82924f375f12db20dd577f2c139a8ff82a2d08159
-size 3093445

elo_results_20240828.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b7da13b5f061a7a5a112e5ca45ff707d6cf6259c8a01b40ea5b77bbd5bd3d5b0
-size 3819732

elo_results_20240904.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:cf6117b1e28bb982e965d20b927685dce98750a82c255b868588b5b2318aaee9
-size 3486555

elo_results_20240915.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:42512fe148c81eeab05961ed64fe446dc7f6ed3703f976fc4c8c2a6a3a3e6bef
-size 3726145

elo_results_20240917.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d15ed36a6e429a9213e9230fbd28de05e7788f758c4282660dd77a4689f98590
-size 3768775

elo_results_20240927.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7c96f9a95fbbf44b8cb46129cc8ea09eef1bb6e43b8ec12d1c8837091d23ee69
-size 3860967

elo_results_20241007.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3891408ad17e52226989783562eb75f784dfb1d0a30dd9051e943b48cd0b117c
-size 3919717

elo_results_20241015.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6daea1d6b24298f32f23c903872aa6c852453de29bf82aa0d2e7aaf5337af715
-size 4058396

elo_results_20241023.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:2e592a6807b4affbc2e203b385ff491743116b4854c9bcd3c3bbf037fba6092f
-size 4374135

elo_results_20241028.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:eb1b06bd57f44c46862bb7c326c39e09fa0e298257d3e8fb96d466a7200fc77d
-size 4417082

elo_results_20241104.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c0129cb6833f0d81f9407489883c588f5712666ecdc35111de725d3502e61a07
-size 4472516

elo_results_20241112.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ab321a67e6831f495d8ac7d6e052f16eae495fb4a82454f8d7bd361bda6f4706
-size 4472047

elo_results_20241113.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:2d0493678e9a9d2d86bdc8adf483824ef731b8efe8a41c5cef1d81bf21f5e7e9
-size 4496349

elo_results_20241120.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:3e5f7d30d02d57f07d0506d385d2007fb7ae4a33a24a2e8e62fc95e9434c565e
-size 4585156

elo_results_20241121.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:610c77a4c4b47f1bb256a3dfa1808d5d32b5d9c819aa5f34b1aae857a959c303
-size 4613615

elo_results_20241122.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5387ed94e44688ff4cacc52644290409dd09580a0c745016c5d473ff9f3368bd
-size 4656304

elo_results_20241201.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e43a1652cfd47ab95459d2a9ac8e5fc8065001f8160dcc8c460c4b1bbda2e58e
-size 4743052

elo_results_20241205.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:651b61a2fdb0f16884975c2d71577f67e311ed0b5675badbe5db2f4b0068d3e5
-size 4757570

elo_results_20241210.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ebcf524b476eba3cc260d9e3f8f8e42d187c56481748645276cd70ea6e1caaa2
-size 4796359

elo_results_20241215.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:af899ac7ceed98b9c127c69fad5b6115bd02e00fe94a2a580734639ef9cdb417
-size 4928341

elo_results_20241218.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9380779e9effeef27f1face0759c2e11075c8b2c4a809f565a9cfd40944bb7a2
-size 4956709

elo_results_20241222.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:030bc6db11694fcf5e025e6aa849d28e61dc4c7bcb66a0b562f7cd2673d9599d
-size 5018497

elo_results_20241230.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:517d699389bdbf7f01ec341306be72ffad2ee36d224a240369f817ae0486db87
-size 5051817

elo_results_20250105.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:609f7b97d9c67d79214c8083a3d8f24f6e0893c3fd45cb079eb2748b7acbc5e8
-size 5979482

elo_results_20250115.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:758c111032a6e355760d5d4074b435174100d43faf1d96ca07e34b76477c56a3
-size 6054581

elo_results_20250119.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:27105cd16c8d3456548c8ab8e93d52abc71634b1e70237173b260cb0bbab93df
-size 6082897

elo_results_20250121.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:81bfdd73af127cf20a9c5effd9871d7d5c5b1dd2ce1e9c2c5076b45feff844df
-size 6111217

elo_results_20250122.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d9e220831e892b7f69e6ac0f4d21c3b091aaf59b0b51c7a19e5618980b07068a
-size 6115326

elo_results_20250124.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b3ea59502af8135c8cd24074d6a1ff5f710553ab46acaab9259a6bbeed535851
-size 6234818

elo_results_20250128.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bdbd9f8bacb67e3e4e829e077b5621fb443f50b6ce0f73ab017607ffccccce1a
-size 6172158

elo_results_20250203.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c591519e3818519bace906c2a79aa9cba7cb09ee5074a7f389def5ff84c1b6aa
-size 6221605

elo_results_20250205.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:53603d4d811e02261964a97bc8cbb46be4380f914388667974ba483e94262e3e
-size 6305480

elo_results_20250206.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8d3e6a23cd1fd5fc826ffbdc3084256e6f375e02b1a2bd950f3ca8da8e38ad17
-size 6330761

elo_results_20250209.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4f4adfd4b26595c489a0aa41dc5357cb97536c9f67f87e6a7da3bfa8ea2b7d57
-size 6334820

elo_results_20250211.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d1aec63fbea7585a4744bbbc95a0ef89503b9121ae0e8527408d3ff7661f388b
-size 6387624

elo_results_20250214.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5134236a0ed3ab0b651e17302749a3dbb0b4c97322722274adf6e42bfa995dc3
-size 6381600

elo_results_20250217.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4cdd16087da85d7b31beaf8ceb0e7787e4b72845203640c83f4d38f2795cd632
-size 6405776