Login to the Hub first

background_inference.py

@@ -5,6 +5,10 @@ import datasets
 import eval_utils
 from constants import DIALECTS_WITH_LABELS
 from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
+from huggingface_hub import login
+
+access_token = os.environ["HF_TOKEN"]
+login(token=access_token)
 
 model_name = sys.argv[1]
 commit_id = sys.argv[2]
@@ -19,16 +23,12 @@ utils.update_model_queue(
 )
 
 try:
-    tokenizer = AutoTokenizer.from_pretrained(
-        model_name, revision=commit_id, access_token=os.environ["HF_TOKEN"]
-    )
+    tokenizer = AutoTokenizer.from_pretrained(model_name, revision=commit_id)
     if inference_function == "prompt_chat_LLM":
-        model = AutoModel.from_pretrained(
-            model_name, revision=commit_id, access_token=os.environ["HF_TOKEN"]
-        )
+        model = AutoModel.from_pretrained(model_name, revision=commit_id)
     else:
         model = AutoModelForSequenceClassification.from_pretrained(
-            model_name, revision=commit_id, access_token=os.environ["HF_TOKEN"]
+            model_name, revision=commit_id
         )
 
     # Load the dataset

leaderboard_info.md

@@ -1,8 +1,8 @@
 ## Test Set Details
-The test set used for evaluation is composed of 1000 sentences geolocated to the 14 most-populated Arab countries (excluding Somalia from which data was scarce). Each sample is annotated by native speakers recruited from 9 different Arab countries, namely: Algeria, Egypt, Iraq, Morocco, Palestine, Sudan, Syria, Tunisia, Yemen.
+The test set used for evaluation is composed of 1000 sentences geolocated to the 14 most-populated Arab countries (excluding Somalia from which data was scarce). Each sample is annotated by native speakers recruited from 11 different Arab countries, namely: Algeria, Egypt, Iraq, Jordan, Morocco, Palestine, Saudi Arabia, Sudan, Syria, Tunisia, Yemen.
 
 ## Evaluation Metrics
-We compute the precision, recall, and F1 scores for each of the 9 countries (treating each label as a binary classification problem). Afterward, 
+We compute the precision, recall, and F1 scores for each of the 11 countries (treating each label as a binary classification problem).
 
 ## Data Access
 If you need to access the single-label training sets, and the multi-label development set, please fill the following form: https://forms.gle/t3QTC6ZqyDJBzAau8

requirements.txt

@@ -5,3 +5,4 @@ pandas
 numpy
 scikit-learn
 tabulate
+huggingface_hub