update with repo links

app.py

@@ -17,8 +17,11 @@ from content import format_warning, format_log, TITLE, INTRODUCTION_TEXT, CHANGE
 
 BALM_TOKEN = os.environ.get("BALM_TOKEN", None)
 
-OWNER="balm"
+OWNER="gaia-benchmark"
+DATA_DATASET = f"{OWNER}/GAIA"
 SUBMISSION_DATASET = f"{OWNER}/submissions"
+RESULTS_DATASET = f"{OWNER}/results"
+LEADERBOARD_PATH = f"{OWNER}/leaderboard"
 
 SPLIT="validation" #Change to test once we are ready to go 
 api = HfApi()
@@ -28,7 +31,8 @@ os.makedirs("scored", exist_ok=True)
 # Display the results
 eval_results = {}
 for level in range(1, 4):
-    eval_results[level] = load_dataset(f"{OWNER}/BALM_ResultsLevel{level}", use_auth_token=BALM_TOKEN, split=SPLIT)
+    eval_results[level] = load_dataset(RESULTS_DATASET, f"2023_level{level}", use_auth_token=BALM_TOKEN, split=SPLIT)
+
 
 eval_dataframe_1 = pd.DataFrame(eval_results[1].remove_columns("mail"))
 eval_dataframe_2 = pd.DataFrame(eval_results[2].remove_columns("mail"))
@@ -37,12 +41,12 @@ eval_dataframe_3 = pd.DataFrame(eval_results[3].remove_columns("mail"))
 # Gold answers
 gold_results = {}
 for level in range(1, 4):
-    level_dataset = load_dataset(f"{OWNER}/BALM", f"2023_level{level}", split=SPLIT, use_auth_token=BALM_TOKEN)
+    level_dataset = load_dataset(DATA_DATASET, f"2023_level{level}", split=SPLIT, use_auth_token=BALM_TOKEN)
     gold_results[level] = {row["task_id"]: row["ground_truth"] for row in level_dataset}
 
 
 def restart_space():
-    api.restart_space(repo_id=f"{OWNER}/BALM_Leaderboard", token=BALM_TOKEN)
+    api.restart_space(repo_id=LEADERBOARD_PATH, token=BALM_TOKEN)
 
 
 COLS = ["Model", "Score ⬆️", "Organisation"]

content.py

@@ -1,9 +1,4 @@
-CHANGELOG_TEXT = f"""
-## [2023-06-02] 
-- Beta internal version of the leaderboard
-"""
-
-TITLE = """<h1 align="center" id="space-title">BALM Leaderboard</h1>"""
+TITLE = """<h1 align="center" id="space-title">GAIA Leaderboard</h1>"""
 
 CANARY_STRING = "" # TODO
 
@@ -11,9 +6,9 @@ INTRODUCTION_TEXT = f"""
 Large language models have seen their potential capabilities increased by several orders of magnitude with the introduction of augmentations, from simple prompting adjustement to actual external tooling (calculators, vision models, ...) or online web retrieval.
 
 To evaluate the next generation of LLMs, we argue for a new kind of benchmark, simple and yet effective to measure actual progress on augmented capabilities,
-We therefore present BALM.
+We therefore present GAIA.
 
-BALM is made of 3 evaluation levels, depending on the added level of tooling and autonomy the model needs.
+GAIA is made of 3 evaluation levels, depending on the added level of tooling and autonomy the model needs.
 We expect the level 1 to be breakable by very good LLMs, and the level 3 to indicate a strong jump in model capabilities.
 
 Each of these levels is divided into two sets: a public dev set, on which people can self report their results, and a private test set, which will be unlocked once public performance passes a threshold on the dev set.
@@ -22,12 +17,12 @@ Please do not repost the public dev set, nor use it in training data for your mo
 """
 
 CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
-CITATION_BUTTON_TEXT = r"""@misc{balm, # TODO
+CITATION_BUTTON_TEXT = r"""@misc{gaia, # TODO
   author = {tbd},
-  title = {Benchmark for Augmented Language Models},
+  title = {General AI Assistant benchamrk},
   year = {2023},
   #publisher = {Hugging Face},
-  #howpublished = "\url{https://huggingface.co/spaces/balm/}"
+  #howpublished = "\url{https://huggingface.co/spaces/gaia-benchmark/}"
 }"""