Update data; Other changes

app.py

@@ -78,7 +78,7 @@ def create_chip_designers_data_center_revenue_plot():
 def create_size_for_performance_plot(category_to_display: str,
                                      parameter_type_to_display: str,
                                      model_to_compare: str) -> (go.Figure, gr.Dropdown, gr.Dropdown):
-    with open('elo_results_20240823.pkl', 'rb') as file:
+    with open('elo_results_20240915.pkl', 'rb') as file:
         elo_results = pickle.load(file)
     categories: list[str] = list(elo_results["text"].keys())
     if category_to_display not in categories:
@@ -136,7 +136,7 @@ def create_size_for_performance_plot(category_to_display: str,
         yaxis_type='log',
         hovermode='x unified',
         xaxis=dict(
-            range=[date(2023, 2, 27), date(2024, 8, 23)],
+            range=[date(2023, 2, 27), date(2024, 9, 15)],
             type='date'
         ),
         height=800
@@ -161,14 +161,12 @@ def create_size_for_performance_plot(category_to_display: str,
 
 with gr.Blocks() as demo:
     with gr.Tab("Finance"):
-        with gr.Tab("Big Five Capex"):
+        with gr.Tab("Big Five Capex") as big_five_capex_tab:
             big_five_capex_plot: gr.Plot = gr.Plot()
-            big_five_capex_button: gr.Button = gr.Button("Show")
-        with gr.Tab("Chip Designers Data Center Revenue"):
+        with gr.Tab("Chip Designers Data Center Revenue") as chip_designers_data_center_revenue_tab:
             chip_designers_data_center_revenue_plot: gr.Plot = gr.Plot()
-            chip_designers_data_center_revenue_button: gr.Button = gr.Button("Show")
-    with gr.Tab("Model Efficiency"):
-        with gr.Tab("Parameters Necessary for Specific Performance Level"):
+    with gr.Tab("Model Efficiency Over Time"):
+        with gr.Tab("Parameters Necessary for Specific Performance Level") as size_for_performance_tab:
             with gr.Row():
                 size_for_performance_category_dropdown: gr.Dropdown = gr.Dropdown(label="Category",
                                                                                   value="full",
@@ -190,24 +188,39 @@ with gr.Blocks() as demo:
             )
         with gr.Tab("API Cost for Specific Performance Level", interactive=False):
             api_cost_for_performance_plot: gr.Plot = gr.Plot()
-            api_cost_for_performance_button: gr.Button = gr.Button("Show")
-    with gr.Tab("AI System Performance", interactive=False):
-        with gr.Tab("SWE-bench"):
-            swe_bench_plot: gr.Plot = gr.Plot()
-            swe_bench_button: gr.Button = gr.Button("Show")
+    with gr.Tab("System Performance Over Time", interactive=False):
+        with gr.Tab("ARC-AGI"):
+            arc_agi_plot: gr.Plot = gr.Plot()
+        with gr.Tab("BigCodeBench"):
+            bigcodebench_plot: gr.Plot = gr.Plot()
+        with gr.Tab("Codeforces"):
+            codeforces_plot: gr.Plot = gr.Plot()
         with gr.Tab("GAIA"):
             gaia_plot: gr.Plot = gr.Plot()
-            gaia_button: gr.Button = gr.Button("Show")
+        with gr.Tab("GPQA"):
+            gpqa_plot: gr.Plot = gr.Plot()
+        with gr.Tab("HumanEval"):
+            humaneval_plot: gr.Plot = gr.Plot()
+        with gr.Tab("LMSYS"):
+            lmsys_plot: gr.Plot = gr.Plot()
+        with gr.Tab("OpenCompass"):
+            opencompass_plot: gr.Plot = gr.Plot()
+        with gr.Tab("SWE-bench"):
+            swe_bench_plot: gr.Plot = gr.Plot()
+        with gr.Tab("Simple Bench"):
+            simple_bench_plot: gr.Plot = gr.Plot()
+        with gr.Tab("WebArena"):
+            webarena_plot: gr.Plot = gr.Plot()
+        with gr.Tab("ZeroEval"):
+            zeroeval_plot: gr.Plot = gr.Plot()
     with gr.Tab("Frontier Language Model Training Runs", interactive=False):
         with gr.Tab("Street Price of GPUs Used"):
             gpu_street_price_plot: gr.Plot = gr.Plot()
-            gpu_street_price_button: gr.Button = gr.Button("Show")
         with gr.Tab("TDP of GPUs Used"):
             tdp_gpus_plot: gr.Plot = gr.Plot()
-            tdp_gpus_button: gr.Button = gr.Button("Show")
-    big_five_capex_button.click(fn=create_big_five_capex_plot, outputs=big_five_capex_plot)
-    chip_designers_data_center_revenue_button.click(fn=create_chip_designers_data_center_revenue_plot,
-                                                    outputs=chip_designers_data_center_revenue_plot)
+    big_five_capex_tab.select(fn=create_big_five_capex_plot, outputs=big_five_capex_plot)
+    chip_designers_data_center_revenue_tab.select(fn=create_chip_designers_data_center_revenue_plot,
+                                                  outputs=chip_designers_data_center_revenue_plot)
     size_for_performance_button.click(fn=create_size_for_performance_plot,
                                       inputs=[size_for_performance_category_dropdown,
                                               size_for_performance_parameter_number_dropdown,

elo_results_20240823.pkl → elo_results_20240915.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ec103aa6bf5d0f02f8bd2c69c8ccfc8f1be1b44c7dc004d967c8d5ce470975b5
-size 3039588
+oid sha256:bdce5fbf7a50d53ce549fd2c6c230627397856c4b62807b0e97a7f2c8554045e
+size 3707205

models.jsonl

@@ -1,12 +1,22 @@
+{"Name": "deepseek-v2.5", "Release Date": "2024-09-05", "Total Parameters": 236, "Active Parameters": 236, "API Cost": 0}
+{"Name": "qwen-plus-0828", "Release Date": "2024-08-28", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "qwen-plus-0828", "Release Date": "2024-08-28", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "gemini-1.5-pro-exp-0827", "Release Date": "2024-08-27", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "gemini-1.5-flash-exp-0827", "Release Date": "2024-08-27", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "gemini-1.5-flash-8b-exp-0827", "Release Date": "2024-08-27", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "chatgpt-4o-latest", "Release Date": "2024-08-25", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "command-r-plus-08-2024", "Release Date": "2024-08-21", "Total Parameters": 104, "Active Parameters": 104, "API Cost": 0}
+{"Name": "command-r-08-2024", "Release Date": "2024-08-19", "Total Parameters": 32, "Active Parameters": 32, "API Cost": 0}
 {"Name": "gemini-1.5-pro-exp-0801", "Release Date": "2024-08-01", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "grok-2-2024-08-13", "Release Date": "2024-08-13", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "gpt-4o-2024-05-13", "Release Date": "2024-05-13", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "gpt-4o-mini-2024-07-18", "Release Date": "2024-07-18", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
+{"Name": "gemma-2-9b-it-simpo", "Release Date": "2024-07-17", "Total Parameters": 9, "Active Parameters": 9, "API Cost": 0}
 {"Name": "claude-3-5-sonnet-20240620", "Release Date": "2024-06-20", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "grok-2-mini-2024-08-13", "Release Date": "2024-08-13", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "gemini-advanced-0514", "Release Date": "2024-05-14", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
-{"Name": "llama-3.1-405b-instruct", "Release Date": "2024-07-23", "Total Parameters": 405, "Active Parameters": 405, "API Cost": 0}
+{"Name": "llama-3.1-405b-instruct-bf16", "Release Date": "2024-07-23", "Total Parameters": 405, "Active Parameters": 405, "API Cost": 0}
+{"Name": "llama-3.1-405b-instruct-fp8", "Release Date": "2024-07-23", "Total Parameters": 405, "Active Parameters": 405, "API Cost": 0}
 {"Name": "gpt-4o-2024-08-06", "Release Date": "2024-08-06", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "gemini-1.5-pro-api-0514", "Release Date": "2024-05-14", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "gemini-1.5-pro-api-0409-preview", "Release Date": "2024-04-09", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
@@ -47,7 +57,7 @@
 {"Name": "qwen1.5-110b-chat", "Release Date": "2024-02-04", "Total Parameters": 110, "Active Parameters": 110, "API Cost": 0}
 {"Name": "mistral-large-2402", "Release Date": "2024-02-24", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "yi-1.5-34b-chat", "Release Date": "2024-05-13", "Total Parameters": 34, "Active Parameters": 34, "API Cost": 0}
-{"Name": "reka-flash-21b-20240226-online", "Release Date": "2024-02-26", "Total Parameters": 21, "Active Parameters": 21, "API Cost": 0}
+{"Name": "reka-flash-21b-20240226-online", "Release Date": "2024-02-26", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "llama-3-8b-instruct", "Release Date": "2024-04-18", "Total Parameters": 8, "Active Parameters": 8, "API Cost": 0}
 {"Name": "claude-1", "Release Date": "2023-03-14", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "command-r", "Release Date": "2024-03-21", "Total Parameters": 35, "Active Parameters": 35, "API Cost": 0}
@@ -84,7 +94,7 @@
 {"Name": "nous-hermes-2-mixtral-8x7b-dpo", "Release Date": "2024-01-11", "Total Parameters": 47, "Active Parameters": 13, "API Cost": 0}
 {"Name": "gemma-1.1-7b-it", "Release Date": "2024-03-26", "Total Parameters": 7, "Active Parameters": 7, "API Cost": 0}
 {"Name": "llama2-70b-steerlm-chat", "Release Date": "2023-11-24", "Total Parameters": 70, "Active Parameters": 70, "API Cost": 0}
-{"Name": "pplx-70b-online", "Release Date": "2023-11-29", "Total Parameters": 70, "Active Parameters": 70, "API Cost": 0}
+{"Name": "pplx-70b-online", "Release Date": "2023-11-29", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "deepseek-llm-67b-chat", "Release Date": "2023-11-29", "Total Parameters": 67, "Active Parameters": 67, "API Cost": 0}
 {"Name": "openchat-3.5", "Release Date": "2023-11-01", "Total Parameters": 7, "Active Parameters": 7, "API Cost": 0}
 {"Name": "openhermes-2.5-mistral-7b", "Release Date": "2023-10-29", "Total Parameters": 7, "Active Parameters": 7, "API Cost": 0}
@@ -99,7 +109,7 @@
 {"Name": "wizardlm-13b", "Release Date": "2023-05-26", "Total Parameters": 13, "Active Parameters": 13, "API Cost": 0}
 {"Name": "zephyr-7b-beta", "Release Date": "2023-10-16", "Total Parameters": 7, "Active Parameters": 7, "API Cost": 0}
 {"Name": "mpt-30b-chat", "Release Date": "2023-04-22", "Total Parameters": 30, "Active Parameters": 30, "API Cost": 0}
-{"Name": "pplx-7b-online", "Release Date": "2023-11-29", "Total Parameters": 7, "Active Parameters": 7, "API Cost": 0}
+{"Name": "pplx-7b-online", "Release Date": "2023-11-29", "Total Parameters": 0, "Active Parameters": 0, "API Cost": 0}
 {"Name": "codellama-70b-instruct", "Release Date": "2024-01-29", "Total Parameters": 70, "Active Parameters": 70, "API Cost": 0}
 {"Name": "codellama-34b-instruct", "Release Date": "2023-08-24", "Total Parameters": 34, "Active Parameters": 34, "API Cost": 0}
 {"Name": "vicuna-13b", "Release Date": "2023-06-22", "Total Parameters": 13, "Active Parameters": 13, "API Cost": 0}