Update monitor & plots (#2506)

Author: merrymercy

docs/commands/leaderboard.md

@@ -24,3 +24,14 @@ scp atlas:/data/lmzheng/FastChat/fastchat/serve/monitor/elo_results_20230905.pkl
 ```
 wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/raw/main/leaderboard_table_20230905.csv
 ```
+
+### Update files on webserver
+```
+DATE=20231002
+
+rm -rf elo_results.pkl leaderboard_table.csv
+wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/resolve/main/elo_results_$DATE.pkl
+wget https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard/resolve/main/leaderboard_table_$DATE.csv
+ln -s leaderboard_table_$DATE.csv leaderboard_table.csv
+ln -s elo_results_$DATE.pkl elo_results.pkl
+```

fastchat/serve/huggingface_api_worker.py

@@ -1,18 +1,19 @@
 """
-A model worker to call huggingface api.
-JSON file format:
+A model worker that calls huggingface inference endpoint.
+
+Register models in a JSON file with the following format:
 {
     "falcon-180b-chat": {
         "model_path": "tiiuae/falcon-180B-chat",
         "api_base": "https://api-inference.huggingface.co/models",
         "token": "hf_xxx",
-        "context_length": 2048
+        "context_length": 2048,
         "model_names": "falcon-180b-chat",
-        "conv_template": null,
+        "conv_template": null
     }
 }
 
-Only "model_path", "api_base", and "token" are necessary, others are optional.
+"model_path", "api_base", "token", and "context_length" are necessary, while others are optional.
 """
 import argparse
 import asyncio
@@ -116,6 +117,9 @@ def __init__(
             f"Connecting with huggingface api {self.model_path} as {self.model_names} on worker {worker_id} ..."
         )
 
+        if not no_register:
+            self.init_heart_beat()
+
     def count_token(self, params):
         # No tokenizer here
         ret = {
@@ -312,7 +316,7 @@ def create_huggingface_api_worker():
         api_base_list.append(model_info[m]["api_base"])
         token_list.append(model_info[m]["token"])
 
-        context_length = model_info[m].get("context_length", 1024)
+        context_length = model_info[m]["context_length"]
         model_names = model_info[m].get("model_names", [m.split("/")[-1]])
         if isinstance(model_names, str):
             model_names = [model_names]

fastchat/serve/monitor/clean_battle_data.py

@@ -44,8 +44,8 @@
 
 def get_log_files(max_num_files=None):
     dates = []
-    for month in [4, 5, 6, 7, 8, 9]:
-        for day in range(1, 32):
+    for month in range(4, 12):
+        for day in range(1, 33):
             dates.append(f"2023-{month:02d}-{day:02d}")
 
     filenames = []

fastchat/serve/monitor/clean_chat_data.py

@@ -28,8 +28,8 @@
 
 def get_log_files(max_num_files=None):
     dates = []
-    for month in [4, 5, 6, 7, 8, 9, 10]:
-        for day in range(1, 32):
+    for month in range(4, 12):
+        for day in range(1, 33):
             dates.append(f"2023-{month:02d}-{day:02d}")
 
     filenames = []

fastchat/serve/monitor/elo_analysis.py

@@ -58,7 +58,7 @@ def get_median_elo_from_bootstrap(bootstrap_df):
     return median
 
 
-def compute_pairwise_win_fraction(battles, model_order):
+def compute_pairwise_win_fraction(battles, model_order, limit_show_number=None):
     # Times each model wins as Model A
     a_win_ptbl = pd.pivot_table(
         battles[battles["winner"] == "model_a"],
@@ -92,6 +92,9 @@ def compute_pairwise_win_fraction(battles, model_order):
         prop_wins = row_beats_col_freq.mean(axis=1).sort_values(ascending=False)
         model_order = list(prop_wins.keys())
 
+    if limit_show_number is not None:
+        model_order = model_order[:limit_show_number]
+
     # Arrange ordering according to proprition of wins
     row_beats_col = row_beats_col_freq.loc[model_order, model_order]
     return row_beats_col
@@ -166,8 +169,10 @@ def visualize_battle_count(battles, model_order):
     return fig
 
 
-def visualize_average_win_rate(battles):
-    row_beats_col_freq = compute_pairwise_win_fraction(battles, None)
+def visualize_average_win_rate(battles, limit_show_number):
+    row_beats_col_freq = compute_pairwise_win_fraction(
+        battles, None, limit_show_number=limit_show_number
+    )
     fig = px.bar(
         row_beats_col_freq.mean(axis=1).sort_values(ascending=False),
         text_auto=".2f",
@@ -180,7 +185,7 @@ def visualize_average_win_rate(battles):
     return fig
 
 
-def visualize_bootstrap_elo_rating(df):
+def visualize_bootstrap_elo_rating(df, limit_show_number):
     bars = (
         pd.DataFrame(
             dict(
@@ -192,6 +197,7 @@ def visualize_bootstrap_elo_rating(df):
         .reset_index(names="model")
         .sort_values("rating", ascending=False)
     )
+    bars = bars[:limit_show_number]
     bars["error_y"] = bars["upper"] - bars["rating"]
     bars["error_y_minus"] = bars["rating"] - bars["lower"]
     bars["rating_rounded"] = np.round(bars["rating"], 2)
@@ -225,12 +231,19 @@ def report_elo_analysis_results(battles_json):
     model_order = list(elo_rating_median.keys())
     model_order.sort(key=lambda k: -elo_rating_median[k])
 
+    limit_show_number = 25  # limit show number to make plots smaller
+    model_order = model_order[:limit_show_number]
+
     # Plots
     leaderboard_table = visualize_leaderboard_table(elo_rating_median)
     win_fraction_heatmap = visualize_pairwise_win_fraction(battles_no_ties, model_order)
     battle_count_heatmap = visualize_battle_count(battles_no_ties, model_order)
-    average_win_rate_bar = visualize_average_win_rate(battles_no_ties)
-    bootstrap_elo_rating = visualize_bootstrap_elo_rating(bootstrap_df)
+    average_win_rate_bar = visualize_average_win_rate(
+        battles_no_ties, limit_show_number
+    )
+    bootstrap_elo_rating = visualize_bootstrap_elo_rating(
+        bootstrap_df, limit_show_number
+    )
 
     last_updated_tstamp = battles["tstamp"].max()
     last_updated_datetime = datetime.datetime.fromtimestamp(

fastchat/serve/monitor/monitor.py

@@ -35,11 +35,11 @@ def make_leaderboard_md(elo_results):
 | [Blog](https://lmsys.org/blog/2023-05-03-arena/) | [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2306.05685) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) | [Twitter](https://twitter.com/lmsysorg) | [Discord](https://discord.gg/HSWAKCrnFx) |
 
 🏆 This leaderboard is based on the following three benchmarks.
-- [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) - a crowdsourced, randomized battle platform. We use 70K+ user votes to compute Elo ratings.
+- [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) - a crowdsourced, randomized battle platform. We use 90K+ user votes to compute Elo ratings.
 - [MT-Bench](https://arxiv.org/abs/2306.05685) - a set of challenging multi-turn questions. We use GPT-4 to grade the model responses.
 - [MMLU](https://arxiv.org/abs/2009.03300) (5-shot) - a test to measure a model's multitask accuracy on 57 tasks.
 
-💻 Code: The Arena Elo ratings are computed by this [notebook]({notebook_url}). The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge). The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval). Higher values are better for all benchmarks. Empty cells mean not available. Last updated: Sept, 2023.
+💻 Code: The Arena Elo ratings are computed by this [notebook]({notebook_url}). The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge). The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval). Higher values are better for all benchmarks. Empty cells mean not available. Last updated: October, 2023.
 """
     return leaderboard_md