Add filters for bad models

background_task.py

@@ -7,6 +7,7 @@ from huggingface_hub import HfApi, Repository
 
 DATASET_REPO_URL = "https://huggingface.co/datasets/huggingface-projects/bot-fight-data"
 DATASET_TEMP_REPO_URL = "https://huggingface.co/datasets/huggingface-projects/temp-match-results"
+FILTER_FILE = "https://huggingface.co/datasets/huggingface-projects/filter-bad-models/raw/main/bad_models.csv"
 ELO_FILENAME = "soccer_elo.csv"
 HISTORY_FILENAME = "soccer_history.csv"
 TEMP_FILENAME = "results.csv"
@@ -103,6 +104,8 @@ class Matchmaking:
                 self.matches["model2"].append(model2.author + "/" + model2.name)
                 self.matches["result"].append(result)
                 self.matches["timestamp"].append(row["timestamp"])
+                model1.games_played += 1
+                model2.games_played += 1
         data_dict = {"model1": [], "model2": [], "timestamp": [], "result": []}
         df = pd.DataFrame(data_dict)
         print(df.head())
@@ -182,11 +185,9 @@ def match(model1, model2):
     model2_id = model2.author + "/" + model2.name
     subprocess.run(["./SoccerTows.x86_64", "-model1", model1_id, "-model2", model2_id, "-nographics", "-batchmode"])
     print(f"Match {model1_id} against {model2_id} ended.")
-    model1.games_played += 1
-    model2.games_played += 1
 
 
-def get_models_list() -> list:
+def get_models_list(filter_bad_models) -> list:
     """
     Get the list of models from the hub and the ELO file.
 
@@ -197,9 +198,14 @@ def get_models_list() -> list:
     data = pd.read_csv(os.path.join(DATASET_REPO_URL, "resolve", "main", ELO_FILENAME))
     models_on_hub = api.list_models(filter=["reinforcement-learning", "ml-agents", "ML-Agents-SoccerTwos"])
     for i, row in data.iterrows():
+        model_id = row["author"] + "/" + row["model"]
+        if model_id in filter_bad_models:
+            continue
         models.append(Model(row["author"], row["model"], row["elo"], row["games_played"]))
-        models_ids.append(row["author"] + "/" + row["model"])
+        models_ids.append(model_id)
     for model in models_on_hub:
+        if model.modelId in filter_bad_models:
+            continue
         author, name = model.modelId.split("/")[0], model.modelId.split("/")[1]
         if model.modelId not in models_ids:
             models.append(Model(author, name))
@@ -228,7 +234,8 @@ def init_matchmaking():
     4. Compute the new ELO rating for each model
     5. Save the results to the hub
     """
-    models = get_models_list()
+    filter_bad_models = pd.read_csv(FILTER_FILE)["model"].tolist()
+    models = get_models_list(filter_bad_models)
     matchmaking = Matchmaking(models)
     matchmaking.run()
     matchmaking.to_csv()