import glob
import json
import math
import os
from dataclasses import dataclass
import dateutil
import numpy as np
from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType, Domains
from src.submission.check_validity import is_model_on_hub
@dataclass
class RankResult:
"""Represents one the overall ranking table
"""
eval_name: str
full_model: str
org: str
model: str
results: dict
license: str = "?"
knowledge_cutoff: str = ""
@classmethod
def init_from_json_dict(self, data):
config = data.get("config")
# Get model and org
model = config.get("model_name")
org = config.get("organization")
license = config.get("license")
knowledge_cutoff = config.get("knowledge_cutoff")
model_results = data.get("results")
# Extract results available in this file (some results are split in several files)
results = {}
for domain in Domains:
domain = domain.value
results[domain.dimension] = model_results.get(domain.dimension).get(domain.metric, None)
return self(
eval_name=f"{org}_{model}",
full_model=f"{org}/{model}",
org=org,
model=model,
results=results,
license=license,
knowledge_cutoff=knowledge_cutoff
)
def to_dict(self):
"""Converts the Eval Result to a dict compatible with our dataframe display"""
# score = 1 / self.results[Domains.dim0.dimension] if self.results[Domains.dim0.dimension] != 0 else 0
# average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
data_dict = {
# "eval_name": self.eval_name, # not a column, just a save name,
# AutoEvalColumn.model.name: make_clickable_model(self.full_model),
AutoEvalColumn.rank.name: None, # placeholder for the rank
AutoEvalColumn.model.name: self.model,
AutoEvalColumn.score.name: self.results[Domains.dim0.value.dimension],
AutoEvalColumn.score_sd.name: None, # placeholder for the score sd
AutoEvalColumn.license.name: self.license,
AutoEvalColumn.organization.name: self.org,
AutoEvalColumn.knowledge_cutoff.name: self.knowledge_cutoff,
# AutoEvalColumn.precision.name: self.precision.value.name,
# AutoEvalColumn.model_type.name: self.model_type.value.name,
# AutoEvalColumn.model_type_symbol.name
# AutoEvalColumn.weight_type.name: self.weight_type.value.name,
# AutoEvalColumn.architecture.name: self.architecture,
# AutoEvalColumn.revision.name: self.revision,
# AutoEvalColumn.average.name: average,
# AutoEvalColumn.likes.name: self.likes,
# AutoEvalColumn.params.name: self.num_params,
# AutoEvalColumn.still_on_hub.name: self.still_on_hub,
}
@dataclass
class ModelResult:
"""Represents one full evaluation. Built from a combination of the result and request file for a given run.
"""
eval_name: str
full_model: str
org: str
model: str
results: dict
license: str = "?"
knowledge_cutoff: str = ""
@classmethod
def init_from_json_dict(self, data):
config = data.get("config")
# Get model and org
model = config.get("model_name")
org = config.get("organization")
license = config.get("license")
knowledge_cutoff = config.get("knowledge_cutoff")
model_results = data.get("results")
new_results = {}
for k, v in model_results.items():
new_v = {}
for kk, vv in v.items():
if vv == 'N/A':
new_v[kk] = None
else:
new_v[kk] = vv
new_results[k] = new_v
# Extract results available in this file (some results are split in several files)
# results = {}
# for domain in Domains:
# domain = domain.value
# results[domain.dimension] = model_results.get(domain.dimension).get(domain.metric, None)
return self(
eval_name=f"{org}_{model}",
full_model=f"{org}/{model}",
org=org,
model=model,
results=new_results,
license=license,
knowledge_cutoff=knowledge_cutoff
)
def to_dict(self):
"""Converts the Eval Result to a dict compatible with our dataframe display"""
data_dict = {
# "eval_name": self.eval_name, # not a column, just a save name,
# AutoEvalColumn.model.name: make_clickable_model(self.full_model),
# AutoEvalColumn.rank.name: None, # placeholder for the rank
AutoEvalColumn.model.name: self.model,
# AutoEvalColumn.score.name: self.results[Domains.dim0.value.dimension],
# AutoEvalColumn.score_sd.name: None, # placeholder for the score sd
# AutoEvalColumn.score_overall.name: float(self.results.get("OVERALL").get("Average Score", None)),
# AutoEvalColumn.score_math_algebra.name: float(self.results.get("Algebra").get("Average Score", None)),
# AutoEvalColumn.score_math_geometry.name: float(self.results.get("Geometry").get("Average Score", None)),
# AutoEvalColumn.score_math_probability.name: float(self.results.get("Probability").get("Average Score", None)),
# AutoEvalColumn.score_reason_logical.name: float(self.results.get("Logical").get("Average Score", None)),
# AutoEvalColumn.score_reason_social.name: float(self.results.get("Social").get("Average Score", None)),
# AutoEvalColumn.sd_overall.name: float(self.results.get("OVERALL").get("Standard Deviation", None)),
# AutoEvalColumn.sd_math_algebra.name: float(self.results.get("Algebra").get("Standard Deviation", None)),
# AutoEvalColumn.sd_math_geometry.name: float(self.results.get("Geometry").get("Standard Deviation", None)),
# AutoEvalColumn.sd_math_probability.name: float(self.results.get("Probability").get("Standard Deviation", None)),
# AutoEvalColumn.sd_reason_logical.name: float(self.results.get("Logical").get("Standard Deviation", None)),
# AutoEvalColumn.sd_reason_social.name: float(self.results.get("Social").get("Standard Deviation", None)),
# AutoEvalColumn.rank_overall.name: int(self.results.get("OVERALL").get("Rank", None)),
# AutoEvalColumn.rank_math_algebra.name: int(self.results.get("Algebra").get("Rank", None)),
# AutoEvalColumn.rank_math_geometry.name: int(self.results.get("Geometry").get("Rank", None)),
# AutoEvalColumn.rank_math_probability.name: int(self.results.get("Probability").get("Rank", None)),
# AutoEvalColumn.rank_reason_logical.name: int(self.results.get("Logical").get("Rank", None)),
# AutoEvalColumn.rank_reason_social.name: int(self.results.get("Social").get("Rank", None)),
AutoEvalColumn.score_overall.name: self.results.get("OVERALL").get("Average Score", None),
AutoEvalColumn.score_math_algebra.name: self.results.get("Algebra").get("Average Score", None),
AutoEvalColumn.score_math_geometry.name: self.results.get("Geometry").get("Average Score", None),
AutoEvalColumn.score_math_probability.name: self.results.get("Probability").get("Average Score", None),
AutoEvalColumn.score_reason_logical.name: self.results.get("Logical").get("Average Score", None),
AutoEvalColumn.score_reason_social.name: self.results.get("Social").get("Average Score", None),
AutoEvalColumn.sd_overall.name: self.results.get("OVERALL").get("Standard Deviation", None),
AutoEvalColumn.sd_math_algebra.name: self.results.get("Algebra").get("Standard Deviation", None),
AutoEvalColumn.sd_math_geometry.name: self.results.get("Geometry").get("Standard Deviation", None),
AutoEvalColumn.sd_math_probability.name: self.results.get("Probability").get("Standard Deviation", None),
AutoEvalColumn.sd_reason_logical.name: self.results.get("Logical").get("Standard Deviation", None),
AutoEvalColumn.sd_reason_social.name: self.results.get("Social").get("Standard Deviation", None),
AutoEvalColumn.rank_overall.name: self.results.get("OVERALL").get("Rank", None),
AutoEvalColumn.rank_math_algebra.name: self.results.get("Algebra").get("Rank", None),
AutoEvalColumn.rank_math_geometry.name: self.results.get("Geometry").get("Rank", None),
AutoEvalColumn.rank_math_probability.name: self.results.get("Probability").get("Rank", None),
AutoEvalColumn.rank_reason_logical.name: self.results.get("Logical").get("Rank", None),
AutoEvalColumn.rank_reason_social.name: self.results.get("Social").get("Rank", None),
AutoEvalColumn.license.name: self.license,
AutoEvalColumn.organization.name: self.org,
AutoEvalColumn.knowledge_cutoff.name: self.knowledge_cutoff,
}
# for task in Tasks:
# data_dict[task.value.col_name] = self.results[task.value.benchmark]
# for domain in Domains:
# data_dict[domain.value.col_name] = self.results[domain.value.dimension]
return data_dict
@dataclass
class EvalResult:
"""Represents one full evaluation. Built from a combination of the result and request file for a given run.
"""
eval_name: str # org_model_precision (uid)
full_model: str # org/model (path on hub)
org: str
model: str
revision: str # commit hash, "" if main
results: dict
precision: Precision = Precision.Unknown
model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
weight_type: WeightType = WeightType.Original # Original or Adapter
architecture: str = "Unknown"
license: str = "?"
likes: int = 0
num_params: int = 0
date: str = "" # submission date of request file
still_on_hub: bool = False
@classmethod
def init_from_json_file(self, json_filepath):
"""Inits the result from the specific model result file"""
with open(json_filepath) as fp:
data = json.load(fp)
config = data.get("config")
# Precision
precision = Precision.from_str(config.get("model_dtype"))
# Get model and org
org_and_model = config.get("model_name", config.get("model_args", None))
org_and_model = org_and_model.split("/", 1)
if len(org_and_model) == 1:
org = None
model = org_and_model[0]
result_key = f"{model}_{precision.value.name}"
else:
org = org_and_model[0]
model = org_and_model[1]
result_key = f"{org}_{model}_{precision.value.name}"
full_model = "/".join(org_and_model)
still_on_hub, _, model_config = is_model_on_hub(
full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
)
architecture = "?"
if model_config is not None:
architectures = getattr(model_config, "architectures", None)
if architectures:
architecture = ";".join(architectures)
# Extract results available in this file (some results are split in several files)
results = {}
for task in Tasks:
task = task.value
# We average all scores of a given metric (not all metrics are present in all files)
accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
if accs.size == 0 or any([acc is None for acc in accs]):
continue
mean_acc = np.mean(accs) * 100.0
results[task.benchmark] = mean_acc
return self(
eval_name=result_key,
full_model=full_model,
org=org,
model=model,
results=results,
precision=precision,
revision= config.get("model_sha", ""),
still_on_hub=still_on_hub,
architecture=architecture
)
def update_with_request_file(self, requests_path):
"""Finds the relevant request file for the current model and updates info with it"""
request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
try:
with open(request_file, "r") as f:
request = json.load(f)
self.model_type = ModelType.from_str(request.get("model_type", ""))
self.weight_type = WeightType[request.get("weight_type", "Original")]
self.license = request.get("license", "?")
self.likes = request.get("likes", 0)
self.num_params = request.get("params", 0)
self.date = request.get("submitted_time", "")
except Exception:
print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
def to_dict(self):
"""Converts the Eval Result to a dict compatible with our dataframe display"""
average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
# print(AutoEvalColumn.precision.name, self.precision.value.name)
data_dict = {
"eval_name": self.eval_name, # not a column, just a save name,
AutoEvalColumn.precision.name: self.precision.value.name,
AutoEvalColumn.model_type.name: self.model_type.value.name,
AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
AutoEvalColumn.weight_type.name: self.weight_type.value.name,
AutoEvalColumn.architecture.name: self.architecture,
AutoEvalColumn.model.name: make_clickable_model(self.full_model),
AutoEvalColumn.revision.name: self.revision,
AutoEvalColumn.average.name: average,
AutoEvalColumn.license.name: self.license,
AutoEvalColumn.likes.name: self.likes,
AutoEvalColumn.params.name: self.num_params,
AutoEvalColumn.still_on_hub.name: self.still_on_hub,
}
for task in Tasks:
data_dict[task.value.col_name] = self.results[task.value.benchmark]
return data_dict
def get_request_file_for_model(requests_path, model_name, precision):
"""Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
request_files = os.path.join(
requests_path,
f"{model_name}_eval_request_*.json",
)
request_files = glob.glob(request_files)
# Select correct request file (precision)
request_file = ""
request_files = sorted(request_files, reverse=True)
for tmp_request_file in request_files:
with open(tmp_request_file, "r") as f:
req_content = json.load(f)
if (
req_content["status"] in ["FINISHED"]
and req_content["precision"] == precision.split(".")[-1]
):
request_file = tmp_request_file
return request_file
def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
"""From the path of the results folder root, extract all needed info for results"""
model_result_filepaths = []
for root, _, files in os.walk(results_path):
# We should only have json files in model results
if len(files) == 0 or any([not f.endswith(".json") for f in files]):
continue
# Sort the files by date
try:
files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
except dateutil.parser._parser.ParserError:
files = [files[-1]]
for file in files:
model_result_filepaths.append(os.path.join(root, file))
eval_results = {}
for model_result_filepath in model_result_filepaths:
# Creation of result
eval_result = EvalResult.init_from_json_file(model_result_filepath)
eval_result.update_with_request_file(requests_path)
# Store results of same eval together
eval_name = eval_result.eval_name
if eval_name in eval_results.keys():
eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
else:
eval_results[eval_name] = eval_result
results = []
for v in eval_results.values():
try:
v.to_dict() # we test if the dict version is complete
results.append(v)
except KeyError: # not all eval values present
continue
return results
def get_raw_model_results(results_path: str) -> list[EvalResult]:
"""From the path of the results folder root, extract all needed info for results"""
try:
with open(results_path) as fp:
data = json.load(fp)
except:
data = eval(open(results_path).read()) # a list of dicts
# print("data", len(data))
# print(data[0])
# {'config': {'model_name': 'ChatGPT-4o-latest (2024-09-03)',
# 'organization': 'OpenAI', 'license': 'Proprietary',
# 'knowledge_cutoff': '2023/10'},
# 'results': {'math-algebra':
# {'Score': 99.19484702, 'Avg Rank': 1.666666667, 'Min Rank': 1, 'Max Rank': 3},
# 'math-probability': {'Score': 100, 'Avg Rank': 1, 'Min Rank': 1, 'Max Rank': 1},
# 'reasoning-logical': {'Avg Rank': 1, 'Min Rank': 1, 'Max Rank': 1},
# 'overall': {'Avg Rank': 2, 'Min Rank': 2, 'Max Rank': 2}}}
eval_results = {}
for result in data:
# Creation of result
eval_result = ModelResult.init_from_json_dict(result)
# print(eval_result)
# ModelResult(eval_name='OpenAI_ChatGPT-4o-latest (2024-09-03)',
# full_model='OpenAI/ChatGPT-4o-latest (2024-09-03)',
# org='OpenAI', model='ChatGPT-4o-latest (2024-09-03)',
# results={'overall': None}, license='Proprietary', knowledge_cutoff='2023/10')
# all_num_results = eval_result.results
# def get_terminal_values(data):
# terminal_values = []
# for key, value in data.items():
# if isinstance(value, dict):
# terminal_values.extend(get_terminal_values(value))
# else:
# terminal_values.append(value)
# return terminal_values
# all_values = get_terminal_values(all_num_results)
# if 'N/A' in all_values:
# continue
eval_name = eval_result.eval_name
eval_results[eval_name] = eval_result
# # Store results of same eval together
# if eval_name in eval_results.keys():
# eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
# else:
# eval_results[eval_name] = eval_result
results = []
for v in eval_results.values():
# print(v.to_dict())
# exit()
# {'eval_name': 'OpenAI_ChatGPT-4o-latest (2024-09-03)',
# 'Model': '<a target="_blank" href="https://huggingface.co/OpenAI/ChatGPT-4o-latest (2024-09-03)"
# style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">OpenAI/ChatGPT-4o-latest (2024-09-03)</a>',
# 'Hub License': 'Proprietary', 'Organization': 'OpenAI', 'Knowledge cutoff': '2023/10', 'Overall': None}
try:
v.to_dict() # we test if the dict version is complete
results.append(v)
except KeyError: # not all eval values present
continue
return results