xeon27
commited on
Commit
·
a319d81
1
Parent(s):
e7a2635
Use dash symbol for markdown
Browse files- src/populate.py +4 -2
src/populate.py
CHANGED
@@ -18,6 +18,8 @@ for task in Tasks:
|
|
18 |
"source": task.value.source,
|
19 |
}
|
20 |
|
|
|
|
|
21 |
|
22 |
def get_inspect_log_url(model_name: str, benchmark_name: str) -> str:
|
23 |
"""Returns the URL to the log file for a given model and benchmark"""
|
@@ -44,12 +46,12 @@ def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchm
|
|
44 |
|
45 |
# # filter out if any of the benchmarks have not been produced
|
46 |
# df = df[has_no_nan_values(df, benchmark_cols)]
|
47 |
-
df = df.fillna(
|
48 |
print(df["GAIA"].head())
|
49 |
|
50 |
# make values clickable and link to log files
|
51 |
for col in benchmark_cols:
|
52 |
-
df[col] = df[[AutoEvalColumn.model.name, col]].apply(lambda x: f"[{x[col]}]({get_inspect_log_url(model_name=x[AutoEvalColumn.model.name].split('>')[1].split('<')[0], benchmark_name=TASK_NAME_INVERSE_MAP[col]['name'])})" if x[col] !=
|
53 |
|
54 |
# # make task names clickable and link to inspect-evals repository - this creates issues later
|
55 |
# df = df.rename(columns={col: f"[{col}]({TASK_NAME_INVERSE_MAP[col]['source']})" for col in benchmark_cols})
|
|
|
18 |
"source": task.value.source,
|
19 |
}
|
20 |
|
21 |
+
EMPTY_SYMBOL = "&ndash"
|
22 |
+
|
23 |
|
24 |
def get_inspect_log_url(model_name: str, benchmark_name: str) -> str:
|
25 |
"""Returns the URL to the log file for a given model and benchmark"""
|
|
|
46 |
|
47 |
# # filter out if any of the benchmarks have not been produced
|
48 |
# df = df[has_no_nan_values(df, benchmark_cols)]
|
49 |
+
df = df.fillna(EMPTY_SYMBOL)
|
50 |
print(df["GAIA"].head())
|
51 |
|
52 |
# make values clickable and link to log files
|
53 |
for col in benchmark_cols:
|
54 |
+
df[col] = df[[AutoEvalColumn.model.name, col]].apply(lambda x: f"[{x[col]}]({get_inspect_log_url(model_name=x[AutoEvalColumn.model.name].split('>')[1].split('<')[0], benchmark_name=TASK_NAME_INVERSE_MAP[col]['name'])})" if x[col] != EMPTY_SYMBOL else x[col], axis=1)
|
55 |
|
56 |
# # make task names clickable and link to inspect-evals repository - this creates issues later
|
57 |
# df = df.rename(columns={col: f"[{col}]({TASK_NAME_INVERSE_MAP[col]['source']})" for col in benchmark_cols})
|