long-code-arena / src /leaderboard_formatting.py
saridormi's picture
Start rendering actual data + minor improvements
9203553
raw
history blame
857 Bytes
from typing import List
COLUMNS_PRETTY = {
"bleu": "BLEU",
"chrf": "ChrF",
"rouge1": "ROUGE-1",
"rouge2": "ROUGE-2",
"rougeL": "ROUGE-L",
"bertscore": "BERTScore",
"bertscore_normalized": "BERTScore (Normalized)",
"model_name": "Model",
"model_availability": "Availability",
"urls": "URLs",
"context_size": "Context Size",
"submitted_by": "Submitted By",
}
METRICS_PER_TASK = {
"commit_message_generation": [
"BLEU",
"ChrF",
"ROUGE-1",
"ROUGE-2",
"ROUGE-L",
"BERTScore",
"BERTScore (Normalized)",
]
}
def get_columns_per_task(task_id: str) -> List[str]:
metrics_per_task = METRICS_PER_TASK[task_id]
return (
["Model Name", "Availability", "Context Size"]
+ metrics_per_task
+ ["Submitted By", "URLs"]
)