import os
from huggingface_hub import CommitOperationAdd, create_commit, RepoUrl
from huggingface_hub import EvalResult, ModelCard
from huggingface_hub.repocard_data import eval_results_to_model_index
import time
from pytablewriter import MarkdownTableWriter
import gradio as gr
from openllm import get_json_format_data, get_datas
import pandas as pd
import traceback
BOT_HF_TOKEN = os.getenv('BOT_HF_TOKEN')
data = get_json_format_data()
finished_models = get_datas(data)
df = pd.DataFrame(finished_models)
desc = """
This is an automated PR created with https://huggingface.co/spaces/eduagarcia-temp/portuguese-leaderboard-results-to-modelcard
The purpose of this PR is to add evaluation results from the Open Portuguese LLM Leaderboard to your model card.
If you encounter any issues, please report them to https://huggingface.co/spaces/eduagarcia-temp/portuguese-leaderboard-results-to-modelcard/discussions
"""
def search(df, value):
result_df = df[df["Model Name"] == value]
return result_df.iloc[0].to_dict() if not result_df.empty else None
def get_details_url(repo):
#author, model = repo.split("/")
return f"https://huggingface.co/datasets/eduagarcia-temp/llm_pt_leaderboard_raw_results/tree/main/{repo}"
def get_query_url(repo):
return f"https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query={repo}"
def get_task_summary(results):
return {
"ENEM":
{"dataset_type":"eduagarcia/enem_challenge",
"dataset_name":"ENEM Challenge (No Images)",
"metric_type":"acc",
"metric_value":results["ENEM"],
"dataset_config": None,
"dataset_split":"train",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 3},
"metric_name":"accuracy"
},
"BLUEX":
{"dataset_type":"eduagarcia-temp/BLUEX_without_images",
"dataset_name":"BLUEX (No Images)",
"metric_type":"acc",
"metric_value":results["BLUEX"],
"dataset_config": None,
"dataset_split":"train",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 3},
"metric_name":"accuracy"
},
"OAB Exams":
{"dataset_type":"eduagarcia/oab_exams",
"dataset_name":"OAB Exams",
"metric_type":"acc",
"metric_value":results["OAB Exams"],
"dataset_config": None,
"dataset_split":"train",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 3},
"metric_name":"accuracy"
},
"ASSIN2 RTE":
{"dataset_type":"assin2",
"dataset_name":"Assin2 RTE",
"metric_type":"f1_macro",
"metric_value":results["ASSIN2 RTE"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 15},
"metric_name":"f1-macro"
},
"ASSIN2 STS":
{"dataset_type":"assin2",
"dataset_name":"Assin2 STS",
"metric_type":"pearson",
"metric_value":results["ASSIN2 STS"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 15},
"metric_name":"pearson"
},
"FAQUAD NLI":
{"dataset_type":"ruanchaves/faquad-nli",
"dataset_name":"FaQuAD NLI",
"metric_type":"f1_macro",
"metric_value":results["FAQUAD NLI"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 15},
"metric_name":"f1-macro"
},
"HateBR":
{"dataset_type":"eduagarcia/portuguese_benchmark",
"dataset_name":"HateBR Binary",
"metric_type":"f1_macro",
"metric_value":results["HateBR"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 25},
"metric_name":"f1-macro"
},
"PT Hate Speech":
{"dataset_type":"eduagarcia/portuguese_benchmark",
"dataset_name":"PT Hate Speech Binary",
"metric_type":"f1_macro",
"metric_value":results["PT Hate Speech"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 25},
"metric_name":"f1-macro"
},
"tweetSentBR":
{"dataset_type":"eduagarcia-temp/tweetsentbr",
"dataset_name":"tweetSentBR",
"metric_type":"f1_macro",
"metric_value":results["tweetSentBR"],
"dataset_config": None,
"dataset_split":"test",
"dataset_revision":None,
"dataset_args":{"num_few_shot": 25},
"metric_name":"f1-macro"
}
}
def get_eval_results(repo):
results = search(df, repo)
task_summary = get_task_summary(results)
md_writer = MarkdownTableWriter()
md_writer.headers = ["Metric", "Value"]
md_writer.value_matrix = [["Average", f"**{results['Average ⬆️']}**"]] + [[v["dataset_name"], v["metric_value"]] for v in task_summary.values()]
text = f"""
# [Open Portuguese LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard)
Detailed results can be found [here]({get_details_url(repo)})
{md_writer.dumps()}
"""
return text
def get_edited_yaml_readme(repo, token: str | None):
card = ModelCard.load(repo, token=token)
results = search(df, repo)
common = {"task_type": 'text-generation', "task_name": 'Text Generation', "source_name": "Open Portuguese LLM Leaderboard", "source_url": get_query_url(repo)}
tasks_results = get_task_summary(results)
if not card.data['eval_results']: # No results reported yet, we initialize the metadata
card.data["model-index"] = eval_results_to_model_index(repo.split('/')[1], [EvalResult(**task, **common) for task in tasks_results.values()])
else: # We add the new evaluations
for task in tasks_results.values():
cur_result = EvalResult(**task, **common)
if any(result.is_equal_except_value(cur_result) for result in card.data['eval_results']):
continue
card.data['eval_results'].append(cur_result)
return str(card)
def commit(repo, pr_number=None, message="Adding Evaluation Results", oauth_token: gr.OAuthToken | None = None): # specify pr number if you want to edit it, don't if you don't want
if oauth_token is None:
gr.Warning("You are not logged in; therefore, the leaderboard-pr-bot will open the pull request instead of you. Click on 'Sign in with Huggingface' to log in.")
token = BOT_HF_TOKEN
elif oauth_token.expires_at < time.time():
raise gr.Error("Token expired. Logout and try again.")
else:
token = oauth_token.token
if repo.startswith("https://huggingface.co/"):
try:
repo = RepoUrl(repo).repo_id
except Exception:
raise gr.Error(f"Not a valid repo id: {str(repo)}")
edited = {"revision": f"refs/pr/{pr_number}"} if pr_number else {"create_pr": True}
try:
try: # check if there is a readme already
readme_text = get_edited_yaml_readme(repo, token=token) + get_eval_results(repo)
except Exception as e:
if "Repo card metadata block was not found." in str(e): # There is no readme
readme_text = get_edited_yaml_readme(repo, token=token)
else:
traceback.print_exc()
print(f"Something went wrong: {e}")
liste = [CommitOperationAdd(path_in_repo="README.md", path_or_fileobj=readme_text.encode())]
commit = (create_commit(repo_id=repo, token=token, operations=liste, commit_message=message, commit_description=desc, repo_type="model", **edited).pr_url)
return commit
except Exception as e:
if "Discussions are disabled for this repo" in str(e):
return "Discussions disabled"
elif "Cannot access gated repo" in str(e):
return "Gated repo"
elif "Repository Not Found" in str(e):
return "Repository Not Found"
else:
return e
if __name__ == "__main__":
print(get_eval_results("Qwen/Qwen1.5-72B-Chat"))