mirror of
https://github.com/NousResearch/atropos.git
synced 2026-04-19 12:57:58 +00:00
703 lines
26 KiB
Python
703 lines
26 KiB
Python
"""
|
|
OlympiadBench Evaluation Environment for Atropos (Generative Mode)
|
|
|
|
This environment evaluates models on OlympiadBench - a benchmark for evaluating
|
|
language models on olympiad-level math and physics problems.
|
|
|
|
Dataset: Hothan/OlympiadBench
|
|
Paper: https://arxiv.org/abs/2402.14008
|
|
|
|
The evaluation follows a generative approach:
|
|
- Models receive challenging olympiad problems (Math or Physics)
|
|
- Expected answers can be numerical, expressions, equations, or intervals
|
|
- Answers should be provided in LaTeX format with \\boxed{answer}
|
|
- Supports thinking mode with <think></think> tags for extended reasoning
|
|
|
|
Note: This implementation supports text-only (TO) problems in both English and Chinese.
|
|
Theorem proving (TP) problems are not included as they require different evaluation.
|
|
"""
|
|
|
|
import asyncio
|
|
import random
|
|
import re
|
|
from typing import Dict, List, Optional, Tuple
|
|
|
|
import wandb
|
|
from datasets import load_dataset
|
|
from eval_helpers import (
|
|
ANSWER_TAG_PATTERN,
|
|
create_system_content,
|
|
extract_thinking_content,
|
|
get_default_thinking_prompt,
|
|
save_eval_results,
|
|
validate_thinking_format,
|
|
)
|
|
from pydantic import Field
|
|
from tqdm.asyncio import tqdm_asyncio
|
|
|
|
from atroposlib.envs.base import (
|
|
APIServerConfig,
|
|
BaseEnv,
|
|
BaseEnvConfig,
|
|
)
|
|
|
|
# Available text-only subsets in OlympiadBench
|
|
AVAILABLE_SUBSETS = [
|
|
"OE_TO_maths_en_COMP", # Open-ended, text-only, maths, English, Competition
|
|
"OE_TO_maths_zh_CEE", # Open-ended, text-only, maths, Chinese, College Entrance Exam
|
|
"OE_TO_maths_zh_COMP", # Open-ended, text-only, maths, Chinese, Competition
|
|
"OE_TO_physics_en_COMP", # Open-ended, text-only, physics, English, Competition
|
|
"OE_TO_physics_zh_CEE", # Open-ended, text-only, physics, Chinese, College Entrance Exam
|
|
]
|
|
|
|
# Answer type descriptions (English)
|
|
ANSWER_TYPE_TEXT_EN = {
|
|
"Numerical": "a numerical value",
|
|
"Expression": "an expression",
|
|
"Equation": "an equation",
|
|
"Interval": "an interval",
|
|
}
|
|
|
|
# Answer type descriptions (Chinese)
|
|
ANSWER_TYPE_TEXT_ZH = {
|
|
"Numerical": "数值",
|
|
"Expression": "表达式",
|
|
"Equation": "方程",
|
|
"Interval": "区间",
|
|
}
|
|
|
|
|
|
def get_answer_type_text(answer_type: str, is_chinese: bool, is_multiple: bool) -> str:
|
|
"""Generate answer type instruction text."""
|
|
if "Need_human_evaluate" in answer_type or "Tuple" in answer_type:
|
|
return ""
|
|
|
|
type_dict = ANSWER_TYPE_TEXT_ZH if is_chinese else ANSWER_TYPE_TEXT_EN
|
|
|
|
# Parse answer type
|
|
single_type = None
|
|
for t in ["Numerical", "Expression", "Equation", "Interval"]:
|
|
if t in answer_type:
|
|
single_type = type_dict[t]
|
|
break
|
|
|
|
if not single_type:
|
|
return ""
|
|
|
|
if is_chinese:
|
|
if is_multiple:
|
|
return f"，题目有多个答案，答案类型均为{single_type}"
|
|
else:
|
|
return f"，答案类型为{single_type}"
|
|
else:
|
|
if is_multiple:
|
|
return f"The problem has multiple answers, each of them should be {single_type}. "
|
|
else:
|
|
return f"The answer should be {single_type}. "
|
|
|
|
|
|
class OlympiadBenchEvalConfig(BaseEnvConfig):
|
|
"""Configuration for OlympiadBench evaluation environment."""
|
|
|
|
# Dataset configuration
|
|
dataset_name: str = Field(
|
|
default="Hothan/OlympiadBench", description="HuggingFace dataset name"
|
|
)
|
|
subset: str = Field(
|
|
default="OE_TO_maths_en_COMP",
|
|
description="Dataset subset (see AVAILABLE_SUBSETS)",
|
|
)
|
|
eval_split: str = Field(
|
|
default="train",
|
|
description="Split to evaluate on (train is the only split available)",
|
|
)
|
|
shuffle_seed: int = Field(default=42, description="Random seed for shuffling")
|
|
|
|
# Generation parameters
|
|
eval_temperature: float = Field(
|
|
default=0.6, description="Temperature for evaluation generation"
|
|
)
|
|
eval_max_tokens: int = Field(
|
|
default=0, description="Max tokens for evaluation (0 = use model default)"
|
|
)
|
|
|
|
# System prompt configuration
|
|
custom_system_prompt: Optional[str] = Field(
|
|
default=None, description="Optional custom system prompt"
|
|
)
|
|
|
|
# Thinking mode configuration
|
|
thinking_mode: bool = Field(
|
|
default=True,
|
|
description="Whether to use thinking mode with <think></think> tags",
|
|
)
|
|
custom_thinking_prompt: Optional[str] = Field(
|
|
default=None, description="Optional custom thinking prompt"
|
|
)
|
|
|
|
# Retry and debug configuration
|
|
max_retries: int = Field(
|
|
default=3, description="Maximum retries for failed API calls"
|
|
)
|
|
retry_delay: float = Field(
|
|
default=1.0, description="Delay between retries in seconds"
|
|
)
|
|
min_response_length: int = Field(
|
|
default=1, description="Minimum response length to consider valid"
|
|
)
|
|
full_debug: bool = Field(default=False, description="Enable full debug output")
|
|
|
|
# Override defaults
|
|
group_size: int = 1
|
|
max_num_workers: int = 1024
|
|
max_eval_workers: int = 256
|
|
max_num_workers_per_node: int = 128
|
|
use_wandb: bool = True
|
|
rollout_server_url: str = "http://localhost:8000"
|
|
total_steps: int = 1
|
|
wandb_name: str = "olympiadbench_eval"
|
|
steps_per_eval: int = 1
|
|
|
|
|
|
class OlympiadBenchEvalEnv(BaseEnv):
|
|
"""
|
|
OlympiadBench Evaluation Environment.
|
|
|
|
Evaluates models on olympiad-level math and physics problems.
|
|
Uses generative evaluation with LaTeX boxed answers.
|
|
"""
|
|
|
|
name = "olympiadbench_eval"
|
|
|
|
def __init__(
|
|
self,
|
|
config: OlympiadBenchEvalConfig,
|
|
server_configs: List[APIServerConfig],
|
|
slurm_job_id: Optional[str] = None,
|
|
testing: bool = False,
|
|
):
|
|
super().__init__(config, server_configs, slurm_job_id, testing)
|
|
self.config: OlympiadBenchEvalConfig = config
|
|
self.eval_items: List[Dict] = []
|
|
self._dataset_loaded = False
|
|
|
|
# Pre-compile regex patterns for answer extraction
|
|
self._boxed_pattern = re.compile(r"\\boxed\{([^{}]*(?:\{[^{}]*\}[^{}]*)*)\}")
|
|
self._answer_tag_pattern = ANSWER_TAG_PATTERN
|
|
|
|
@classmethod
|
|
def config_cls(cls) -> type:
|
|
return OlympiadBenchEvalConfig
|
|
|
|
async def setup(self) -> None:
|
|
"""Initialize the environment and load the dataset."""
|
|
await super().setup()
|
|
|
|
if not self._dataset_loaded:
|
|
await self._load_dataset()
|
|
|
|
print("\nOlympiadBench Evaluation Setup (Generative Mode):")
|
|
print(f" Dataset: {self.config.dataset_name}")
|
|
print(f" Subset: {self.config.subset}")
|
|
print(f" Evaluation split: {self.config.eval_split}")
|
|
print(f" Thinking mode: {self.config.thinking_mode}")
|
|
if self.config.thinking_mode:
|
|
thinking_prompt = get_default_thinking_prompt(
|
|
self.config.custom_thinking_prompt
|
|
)
|
|
print(f" Thinking prompt: {thinking_prompt[:80]}...")
|
|
print(f" Loaded {len(self.eval_items)} evaluation items")
|
|
|
|
async def _load_dataset(self) -> None:
|
|
"""Load and process the OlympiadBench dataset."""
|
|
print(
|
|
f"Loading OlympiadBench dataset: {self.config.dataset_name}/{self.config.subset}..."
|
|
)
|
|
|
|
if self.config.subset not in AVAILABLE_SUBSETS:
|
|
print(
|
|
f"Warning: Subset '{self.config.subset}' may not be text-only. Available text-only subsets: {AVAILABLE_SUBSETS}" # noqa: E501
|
|
)
|
|
|
|
try:
|
|
dataset = load_dataset(
|
|
self.config.dataset_name, self.config.subset, trust_remote_code=True
|
|
)
|
|
except Exception as e:
|
|
print(f"Error loading dataset: {e}")
|
|
raise
|
|
|
|
if self.config.eval_split not in dataset:
|
|
available_splits = list(dataset.keys())
|
|
raise ValueError(
|
|
f"Split '{self.config.eval_split}' not found. Available: {available_splits}"
|
|
)
|
|
|
|
split_data = dataset[self.config.eval_split]
|
|
|
|
# Parse subset info
|
|
is_chinese = "_zh_" in self.config.subset
|
|
is_math = "maths" in self.config.subset or "Math" in self.config.subset
|
|
|
|
# Process items
|
|
self.eval_items = []
|
|
for idx, item in enumerate(split_data):
|
|
question = item.get("question", "")
|
|
final_answer = item.get("final_answer", "")
|
|
|
|
if not question or not final_answer:
|
|
continue
|
|
|
|
# Get metadata
|
|
subject = item.get("subject", "Math" if is_math else "Physics")
|
|
language = item.get("language", "Chinese" if is_chinese else "English")
|
|
answer_type = item.get("answer_type", "")
|
|
is_multiple = item.get("is_multiple_answer", False)
|
|
unit = item.get("unit", "")
|
|
|
|
self.eval_items.append(
|
|
{
|
|
"id": str(idx),
|
|
"question": question,
|
|
"answer": final_answer,
|
|
"subject": subject,
|
|
"language": language,
|
|
"answer_type": answer_type,
|
|
"is_multiple_answer": is_multiple,
|
|
"unit": unit,
|
|
"is_chinese": "Chinese" in language,
|
|
}
|
|
)
|
|
|
|
# Shuffle with seed
|
|
random.seed(self.config.shuffle_seed)
|
|
random.shuffle(self.eval_items)
|
|
|
|
self._dataset_loaded = True
|
|
print(f"Loaded {len(self.eval_items)} evaluation items")
|
|
|
|
def _format_prompt(self, item: Dict) -> str:
|
|
"""Format the problem into a prompt with appropriate instructions."""
|
|
is_chinese = item.get("is_chinese", False)
|
|
is_math = "Math" in item.get("subject", "Math")
|
|
is_multiple = item.get("is_multiple_answer", False)
|
|
answer_type = item.get("answer_type", "")
|
|
unit = item.get("unit", "")
|
|
|
|
if is_chinese:
|
|
subject_content = "数学" if is_math else "物理"
|
|
answer_type_text = get_answer_type_text(
|
|
answer_type, is_chinese=True, is_multiple=is_multiple
|
|
)
|
|
|
|
if is_multiple:
|
|
multiple_answer_text = "\\boxed{用英文逗号连接的多个答案}"
|
|
else:
|
|
multiple_answer_text = "\\boxed{答案}"
|
|
|
|
unit_text = ""
|
|
if unit:
|
|
multiple_answer_text += "(单位)"
|
|
unit_text = "，注意答案的单位不要放在\\boxed{}中"
|
|
|
|
instruction = f"以下是{subject_content}竞赛中的解答题{answer_type_text}。请根据题目的要求和所提供的信息计算得出答案。解答过程和结果中使用的变量和公式请使用LaTeX格式表示。" # noqa: E501
|
|
instruction += f"\n\n请将你的最终答案放在<answer></answer>标签中，格式为{multiple_answer_text}{unit_text}。"
|
|
instruction += "\n\n示例格式:\n<answer>\\boxed{42}</answer>"
|
|
else:
|
|
subject = "Math" if is_math else "Physics"
|
|
answer_type_text = get_answer_type_text(
|
|
answer_type, is_chinese=False, is_multiple=is_multiple
|
|
)
|
|
|
|
if is_multiple:
|
|
multiple_answer_text = "\\boxed{multiple answers connected with commas}"
|
|
else:
|
|
multiple_answer_text = "\\boxed{answer}"
|
|
|
|
unit_text = ""
|
|
if unit:
|
|
multiple_answer_text += "(unit)"
|
|
unit_text = ", note that the unit of the answer should not be included in \\boxed{}"
|
|
|
|
instruction = f"The following is an open-ended problem from an International {subject} competition. {answer_type_text}Please calculate the answer according to the given requirements and the information provided. Please use LaTeX format to represent the variables and formulas used in the solution process and results." # noqa: E501
|
|
instruction += f"\n\nProvide your final answer within <answer></answer> tags in the format {multiple_answer_text}{unit_text}." # noqa: E501
|
|
instruction += "\n\nExample format:\n<answer>\\boxed{42}</answer>"
|
|
|
|
return f"{instruction}\n\n{item['question']}"
|
|
|
|
def _create_system_content(self) -> str:
|
|
"""Create system message content based on thinking mode."""
|
|
return (
|
|
create_system_content(
|
|
self.config.thinking_mode,
|
|
self.config.custom_thinking_prompt,
|
|
self.config.custom_system_prompt,
|
|
)
|
|
or ""
|
|
)
|
|
|
|
def _normalize_answer(self, answer: str) -> str:
|
|
"""Normalize an answer for comparison."""
|
|
if not answer:
|
|
return ""
|
|
|
|
# Remove outer whitespace
|
|
normalized = answer.strip()
|
|
|
|
# Remove \boxed{} wrapper if present
|
|
boxed_match = self._boxed_pattern.search(normalized)
|
|
if boxed_match:
|
|
normalized = boxed_match.group(1)
|
|
|
|
# Normalize whitespace
|
|
normalized = " ".join(normalized.split())
|
|
|
|
# Remove common LaTeX spacing commands
|
|
normalized = re.sub(r"\\[,;:!]", "", normalized)
|
|
normalized = re.sub(r"\\quad|\\qquad", " ", normalized)
|
|
|
|
# Normalize common math equivalents
|
|
normalized = normalized.replace("\\times", "*")
|
|
normalized = normalized.replace("\\cdot", "*")
|
|
normalized = normalized.replace("\\div", "/")
|
|
|
|
return normalized.strip()
|
|
|
|
def _check_match(self, predicted: str, gold: str) -> Tuple[bool, str]:
|
|
"""
|
|
Check if the predicted answer matches the gold answer.
|
|
|
|
Handles multiple answers (comma-separated) and various formats.
|
|
"""
|
|
pred_norm = self._normalize_answer(predicted)
|
|
gold_norm = self._normalize_answer(gold)
|
|
|
|
if not pred_norm:
|
|
return False, "empty_prediction"
|
|
|
|
# Exact match
|
|
if pred_norm == gold_norm:
|
|
return True, "exact"
|
|
|
|
# Case-insensitive match
|
|
if pred_norm.lower() == gold_norm.lower():
|
|
return True, "case_insensitive"
|
|
|
|
# Try numeric comparison
|
|
try:
|
|
pred_num = float(pred_norm.replace(",", ""))
|
|
gold_num = float(gold_norm.replace(",", ""))
|
|
if abs(pred_num - gold_num) < 1e-9:
|
|
return True, "numeric_exact"
|
|
# Allow small relative error for floating point
|
|
if gold_num != 0 and abs(pred_num - gold_num) / abs(gold_num) < 1e-6:
|
|
return True, "numeric_approx"
|
|
except (ValueError, TypeError):
|
|
pass
|
|
|
|
# Handle multiple answers (comma-separated)
|
|
if "," in gold_norm:
|
|
gold_parts = set(p.strip() for p in gold_norm.split(","))
|
|
pred_parts = set(p.strip() for p in pred_norm.split(","))
|
|
if gold_parts == pred_parts:
|
|
return True, "multi_answer_set"
|
|
|
|
# Containment check (gold in prediction)
|
|
if gold_norm and gold_norm in pred_norm:
|
|
return True, "gold_contained"
|
|
|
|
return False, "no_match"
|
|
|
|
def _extract_answer(
|
|
self, response: str, debug: bool = False
|
|
) -> Tuple[Optional[str], str]:
|
|
"""
|
|
Extract the answer from the response.
|
|
|
|
Looks for <answer></answer> tags first, then \\boxed{}.
|
|
"""
|
|
if not response:
|
|
return None, "empty_response"
|
|
|
|
# Try <answer></answer> tags first
|
|
answer_tag_match = self._answer_tag_pattern.search(response)
|
|
if answer_tag_match:
|
|
answer_content = answer_tag_match.group(1).strip()
|
|
if answer_content:
|
|
# Check if there's a boxed inside
|
|
boxed_match = self._boxed_pattern.search(answer_content)
|
|
if boxed_match:
|
|
extracted = boxed_match.group(1)
|
|
if debug:
|
|
print(
|
|
f" Extracted '{extracted}' from boxed inside answer tag"
|
|
)
|
|
return extracted, "answer_tag_boxed"
|
|
else:
|
|
if debug:
|
|
print(f" Extracted '{answer_content}' from answer tag")
|
|
return answer_content, "answer_tag"
|
|
|
|
# Fallback: Look for \boxed{} anywhere in response
|
|
boxed_matches = self._boxed_pattern.findall(response)
|
|
if boxed_matches:
|
|
# Take the last boxed answer (most likely to be the final answer)
|
|
extracted = boxed_matches[-1]
|
|
if debug:
|
|
print(f" Extracted '{extracted}' from boxed fallback")
|
|
return extracted, "boxed_fallback"
|
|
|
|
# Last resort: Look for "answer is X" patterns
|
|
patterns = [
|
|
r"(?:the\s+)?(?:final\s+)?answer\s+is\s*:?\s*(.+?)(?:\n|$)",
|
|
r"(?:so\s+)?the\s+answer\s+is\s*:?\s*(.+?)(?:\n|$)",
|
|
r"=\s*([^\n=]+?)(?:\n|$)", # Last equation result
|
|
]
|
|
|
|
for pattern in patterns:
|
|
match = re.search(pattern, response, re.IGNORECASE)
|
|
if match:
|
|
answer = match.group(1).strip()
|
|
if answer and len(answer) < 100: # Sanity check length
|
|
if debug:
|
|
print(f" Extracted '{answer}' from pattern fallback")
|
|
return answer, "pattern_fallback"
|
|
|
|
return None, "no_match"
|
|
|
|
async def rollout_and_score_eval(
|
|
self,
|
|
item: Dict,
|
|
server: APIServerConfig,
|
|
) -> Optional[Dict]:
|
|
"""Run evaluation on a single item and return the result."""
|
|
prompt = self._format_prompt(item)
|
|
system_content = self._create_system_content()
|
|
|
|
messages = []
|
|
if system_content:
|
|
messages.append({"role": "system", "content": system_content})
|
|
messages.append({"role": "user", "content": prompt})
|
|
|
|
# Build API call parameters
|
|
kwargs = {
|
|
"model": server.model_name,
|
|
"messages": messages,
|
|
"temperature": self.config.eval_temperature,
|
|
}
|
|
if self.config.eval_max_tokens > 0:
|
|
kwargs["max_tokens"] = self.config.eval_max_tokens
|
|
|
|
response_text = ""
|
|
for attempt in range(self.config.max_retries):
|
|
try:
|
|
response = await self.server.chat_completion(**kwargs)
|
|
response_text = response.choices[0].message.content or ""
|
|
|
|
if len(response_text) >= self.config.min_response_length:
|
|
break
|
|
|
|
except Exception as e:
|
|
if self.config.full_debug:
|
|
print(f" API error (attempt {attempt + 1}): {e}")
|
|
if attempt < self.config.max_retries - 1:
|
|
await asyncio.sleep(self.config.retry_delay)
|
|
continue
|
|
|
|
if not response_text:
|
|
return None
|
|
|
|
# Validate thinking format and extract content after </think>
|
|
is_valid_format, content_for_extraction = validate_thinking_format(
|
|
response_text, self.config.thinking_mode
|
|
)
|
|
|
|
# Extract thinking content if present
|
|
thinking_content = (
|
|
extract_thinking_content(response_text)
|
|
if self.config.thinking_mode
|
|
else None
|
|
)
|
|
|
|
# Extract answer
|
|
extracted_answer, extraction_method = self._extract_answer(
|
|
content_for_extraction, debug=self.config.full_debug
|
|
)
|
|
|
|
# Check match
|
|
gold_answer = item["answer"]
|
|
is_correct, match_type = (
|
|
self._check_match(extracted_answer, gold_answer)
|
|
if extracted_answer
|
|
else (False, "no_extraction")
|
|
)
|
|
|
|
if self.config.full_debug:
|
|
print(f"\n--- Item: {item['id']} ---")
|
|
print(f"Question: {item['question'][:100]}...")
|
|
print(f"Gold answer: {gold_answer}")
|
|
print(f"Extracted: {extracted_answer}")
|
|
print(f"Match: {is_correct} ({match_type})")
|
|
|
|
return {
|
|
"item_id": item["id"],
|
|
"question": item["question"],
|
|
"subject": item.get("subject", ""),
|
|
"language": item.get("language", ""),
|
|
"answer_type": item.get("answer_type", ""),
|
|
"gold_answer": gold_answer,
|
|
"extracted_answer": extracted_answer,
|
|
"extraction_method": extraction_method,
|
|
"is_correct": is_correct,
|
|
"match_type": match_type,
|
|
"format_valid": is_valid_format,
|
|
"response": response_text,
|
|
"thinking_content": thinking_content,
|
|
"has_thinking": thinking_content is not None,
|
|
}
|
|
|
|
async def evaluate(self, *args, **kwargs) -> Dict:
|
|
"""Run the full OlympiadBench evaluation."""
|
|
print(f"\n{'='*60}")
|
|
print("Starting OlympiadBench Evaluation (Generative Mode)")
|
|
print(f"{'='*60}")
|
|
print(f" Subset: {self.config.subset}")
|
|
print(f" Total questions: {len(self.eval_items)}")
|
|
print(f" Thinking mode: {self.config.thinking_mode}")
|
|
print(f"{'='*60}\n")
|
|
|
|
# Create evaluation tasks
|
|
async def eval_task(item):
|
|
return await self.rollout_and_score_eval(item, self.server_configs[0])
|
|
|
|
tasks = [eval_task(item) for item in self.eval_items]
|
|
|
|
# Run with progress bar
|
|
results = await tqdm_asyncio.gather(*tasks, desc="Evaluating OlympiadBench")
|
|
|
|
# Filter out failed results
|
|
valid_results = [r for r in results if r is not None]
|
|
|
|
if not valid_results:
|
|
print("Warning: No valid evaluation results obtained")
|
|
return {"error": "No valid results", "accuracy": 0.0}
|
|
|
|
# Calculate metrics
|
|
total = len(valid_results)
|
|
correct = sum(1 for r in valid_results if r["is_correct"])
|
|
accuracy = correct / total if total > 0 else 0.0
|
|
|
|
# Calculate per-subject metrics
|
|
subject_metrics = {}
|
|
for r in valid_results:
|
|
subject = r.get("subject", "unknown")
|
|
if subject not in subject_metrics:
|
|
subject_metrics[subject] = {"total": 0, "correct": 0}
|
|
subject_metrics[subject]["total"] += 1
|
|
if r["is_correct"]:
|
|
subject_metrics[subject]["correct"] += 1
|
|
|
|
for subject in subject_metrics:
|
|
s_total = subject_metrics[subject]["total"]
|
|
s_correct = subject_metrics[subject]["correct"]
|
|
subject_metrics[subject]["accuracy"] = (
|
|
s_correct / s_total if s_total > 0 else 0.0
|
|
)
|
|
|
|
# Calculate per-answer-type metrics
|
|
type_metrics = {}
|
|
for r in valid_results:
|
|
atype = r.get("answer_type", "unknown") or "unknown"
|
|
if atype not in type_metrics:
|
|
type_metrics[atype] = {"total": 0, "correct": 0}
|
|
type_metrics[atype]["total"] += 1
|
|
if r["is_correct"]:
|
|
type_metrics[atype]["correct"] += 1
|
|
|
|
for atype in type_metrics:
|
|
t_total = type_metrics[atype]["total"]
|
|
t_correct = type_metrics[atype]["correct"]
|
|
type_metrics[atype]["accuracy"] = (
|
|
t_correct / t_total if t_total > 0 else 0.0
|
|
)
|
|
|
|
# Format compliance and thinking utilization
|
|
format_valid = sum(1 for r in valid_results if r.get("format_valid", True))
|
|
has_thinking = sum(1 for r in valid_results if r.get("has_thinking", False))
|
|
|
|
metrics = {
|
|
"accuracy": accuracy,
|
|
"total_evaluated": total,
|
|
"total_correct": correct,
|
|
"subset": self.config.subset,
|
|
"format_compliance_rate": format_valid / total if total > 0 else 0.0,
|
|
"thinking_utilization_rate": has_thinking / total if total > 0 else 0.0,
|
|
"subject_metrics": subject_metrics,
|
|
"answer_type_metrics": type_metrics,
|
|
}
|
|
|
|
print(f"\n{'='*60}")
|
|
print("OlympiadBench Evaluation Results")
|
|
print(f"{'='*60}")
|
|
print(f" Subset: {self.config.subset}")
|
|
print(f" Overall Accuracy: {accuracy:.2%} ({correct}/{total})")
|
|
print(f" Format Compliance: {format_valid / total:.2%}")
|
|
if self.config.thinking_mode:
|
|
print(f" Thinking Utilization: {has_thinking / total:.2%}")
|
|
if subject_metrics:
|
|
print("\n Per-Subject Breakdown:")
|
|
for subject, data in subject_metrics.items():
|
|
print(
|
|
f" {subject}: {data['accuracy']:.2%} ({data['correct']}/{data['total']})"
|
|
)
|
|
print(f"{'='*60}\n")
|
|
|
|
# Save results
|
|
if self.config.data_dir_to_save_evals:
|
|
self._save_results(metrics, valid_results)
|
|
|
|
return metrics
|
|
|
|
def _save_results(self, metrics: Dict, results: List[Dict]) -> None:
|
|
"""Save evaluation results to disk."""
|
|
save_eval_results(self.config.data_dir_to_save_evals, metrics, results)
|
|
|
|
async def wandb_log(self, metrics: Dict, step: int = 0) -> None:
|
|
"""Log metrics to Weights & Biases."""
|
|
if not self.config.use_wandb:
|
|
return
|
|
|
|
log_dict = {
|
|
"olympiadbench/accuracy": metrics.get("accuracy", 0),
|
|
"olympiadbench/total_evaluated": metrics.get("total_evaluated", 0),
|
|
"olympiadbench/format_compliance_rate": metrics.get(
|
|
"format_compliance_rate", 0
|
|
),
|
|
"olympiadbench/thinking_utilization_rate": metrics.get(
|
|
"thinking_utilization_rate", 0
|
|
),
|
|
}
|
|
|
|
# Log per-subject accuracies
|
|
for subject, data in metrics.get("subject_metrics", {}).items():
|
|
safe_name = subject.replace(" ", "_")[:30]
|
|
log_dict[f"olympiadbench/accuracy_{safe_name}"] = data.get("accuracy", 0)
|
|
|
|
wandb.log(log_dict, step=step)
|
|
|
|
# Required abstract method implementations
|
|
async def get_next_item(self) -> Optional[Dict]:
|
|
"""Not used in evaluation mode."""
|
|
return None
|
|
|
|
async def collect_trajectories(self, item) -> List:
|
|
"""Not used in evaluation mode."""
|
|
return []
|
|
|
|
async def score(self, rollout_group_data) -> Optional[List]:
|
|
"""Not used in evaluation mode."""
|
|
return None
|
|
|
|
|
|
if __name__ == "__main__":
|
|
OlympiadBenchEvalEnv.cli()
|