atropos/environments/eval_environments/bbh_eval.py
2025-12-24 23:36:36 +00:00

607 lines
20 KiB
Python

"""
BigBench Hard (BBH) Evaluation Environment for Atropos (Generative Mode)
This environment evaluates models on BigBench Hard - a collection of 23 challenging
tasks from the BIG-Bench benchmark that are particularly difficult for language models.
Dataset: lighteval/bbh
Paper: https://arxiv.org/abs/2210.09261
The evaluation follows a generative approach:
- Models receive challenging reasoning problems
- All tasks are multiple choice (MCQA) with varying number of choices
- Answers are extracted from <answer></answer> tags
- Supports thinking mode with <think></think> tags for extended reasoning
Available subsets include: causal_judgment, date_understanding, disambiguation_qa,
geometric_shapes, logical_deduction (3/5/7 objects), movie_recommendation,
navigate, reasoning_about_colored_objects, ruin_names, salient_translation_error_detection,
snarks, sports_understanding, temporal_sequences, tracking_shuffled_objects (3/5/7 objects).
"""
import asyncio
import random
from string import ascii_uppercase
from typing import Dict, List, Optional, Tuple
import wandb
from datasets import load_dataset
from eval_helpers import (
build_mcqa_fallback_patterns,
create_system_content,
extract_letter_from_answer_tag,
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,
)
# All available BBH subsets
BBH_SUBSETS = [
"causal_judgement",
"date_understanding",
"disambiguation_qa",
"geometric_shapes",
"logical_deduction_five_objects",
"logical_deduction_seven_objects",
"logical_deduction_three_objects",
"movie_recommendation",
"navigate",
"reasoning_about_colored_objects",
"ruin_names",
"salient_translation_error_detection",
"snarks",
"sports_understanding",
"temporal_sequences",
"tracking_shuffled_objects_five_objects",
"tracking_shuffled_objects_seven_objects",
"tracking_shuffled_objects_three_objects",
]
def format_bbh_prompt(item: Dict) -> Tuple[str, List[str], int]:
"""
Format a BBH item into a prompt.
Args:
item: The dataset item
Returns:
Tuple of (prompt_text, choices_list, gold_index)
"""
# Build the query
task_prefix = item.get("task_prefix", "")
input_prefix = item.get("example_input_prefix", "\nQuestion: ")
input_text = item.get("input", "")
choice_prefix = item.get("choice_prefix", "\n Choices: ")
# Note: output_prefix from item.get("example_output_prefix") is not used in generative mode
choices = item.get("choices", [])
target_idx = item.get("target_idx", 0)
# Build choice text
num_choices = len(choices)
valid_letters = list(ascii_uppercase[:num_choices])
choice_text = ""
for i, (letter, choice) in enumerate(zip(valid_letters, choices)):
choice_text += f"\n{letter}. {choice}"
# Add answer tag instruction
valid_letters_str = (
", ".join(valid_letters[:-1]) + f", or {valid_letters[-1]}"
if len(valid_letters) > 1
else valid_letters[0]
)
query = f"""Answer the following question. Think step by step before answering.
Provide your final answer within <answer></answer> tags, containing only the letter ({valid_letters_str}).
Example format:
<answer>A</answer>
"""
# Add task-specific content
if task_prefix:
query += task_prefix
query += input_prefix
query += input_text
query += choice_prefix
query += choice_text
return query, choices, target_idx
class BBHEvalConfig(BaseEnvConfig):
"""Configuration for BigBench Hard evaluation environment."""
# Dataset configuration
dataset_name: str = Field(
default="lighteval/bbh", description="HuggingFace dataset name"
)
subset: str = Field(
default="all",
description="Subset to evaluate ('all' for all subsets, or specific subset name)",
)
eval_split: str = Field(
default="train",
description="Split to evaluate on (train is typically the only available split)",
)
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 = "bbh_eval"
steps_per_eval: int = 1
class BBHEvalEnv(BaseEnv):
"""
BigBench Hard (BBH) Evaluation Environment.
Evaluates models on challenging reasoning tasks from the BIG-Bench benchmark.
All tasks are multiple choice with answer extraction from <answer></answer> tags.
"""
name = "bbh_eval"
def __init__(
self,
config: BBHEvalConfig,
server_configs: List[APIServerConfig],
slurm_job_id: Optional[str] = None,
testing: bool = False,
):
super().__init__(config, server_configs, slurm_job_id, testing)
self.config: BBHEvalConfig = config
self.eval_items: List[Dict] = []
self._dataset_loaded = False
@classmethod
def config_cls(cls) -> type:
return BBHEvalConfig
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("\nBBH 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 BBH dataset."""
# Determine which subsets to load
if self.config.subset.lower() == "all":
subsets_to_load = BBH_SUBSETS
else:
if self.config.subset not in BBH_SUBSETS:
print(
f"Warning: Subset '{self.config.subset}' may not exist. Available: {BBH_SUBSETS}"
)
subsets_to_load = [self.config.subset]
self.eval_items = []
for subset in subsets_to_load:
print(f"Loading BBH subset: {subset}...")
try:
dataset = load_dataset(
self.config.dataset_name, subset, trust_remote_code=True
)
except Exception as e:
print(f" Error loading subset '{subset}': {e}")
continue
if self.config.eval_split not in dataset:
available_splits = list(dataset.keys())
print(
f" Split '{self.config.eval_split}' not found for {subset}. Available: {available_splits}"
)
continue
split_data = dataset[self.config.eval_split]
# Process items
for idx, item in enumerate(split_data):
# Skip items without choices
choices = item.get("choices", [])
if not choices:
continue
self.eval_items.append(
{
"id": f"{subset}_{idx}",
"subset": subset,
"raw_item": item,
"choices": choices,
"target_idx": item.get("target_idx", 0),
"input": item.get("input", ""),
}
)
print(
f" Loaded {len([i for i in self.eval_items if i['subset'] == subset])} items from {subset}"
)
# Shuffle with seed
random.seed(self.config.shuffle_seed)
random.shuffle(self.eval_items)
self._dataset_loaded = True
print(
f"Total: Loaded {len(self.eval_items)} evaluation items from {len(subsets_to_load)} subsets"
)
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 _extract_answer(
self, response: str, num_choices: int, choices: List[str], debug: bool = False
) -> Tuple[Optional[str], str]:
"""
Extract the letter answer from the response.
Args:
response: The model's response (content after </think> in thinking mode)
num_choices: Number of valid choices
choices: List of choice texts
debug: Whether to print debug information
Returns:
Tuple of (extracted_letter or None, extraction_method)
"""
if not response:
return None, "empty_response"
valid_letters = set(ascii_uppercase[:num_choices])
# PRIMARY: Try <answer></answer> tags
letter, method = extract_letter_from_answer_tag(
response, valid_letters, debug=debug, choices=choices
)
if letter:
return letter, method
# FALLBACK: Use regex patterns
fallback_patterns = build_mcqa_fallback_patterns(num_choices)
for priority, pattern, method_name in fallback_patterns:
matches = pattern.findall(response)
if matches:
# Get the last match for answer patterns
match = (
matches[-1]
if method_name
in [
"final_answer_is",
"the_answer_is",
"answer_colon",
"answer_space",
]
else matches[0]
)
if isinstance(match, tuple):
match = match[0]
extracted = match.strip("()").upper()
if extracted in valid_letters:
if debug:
print(
f" Extracted '{extracted}' using fallback '{method_name}'"
)
return extracted, f"fallback_{method_name}"
# Last resort: find any valid letter (take the last one)
for letter in reversed(list(valid_letters)):
if letter in response.upper():
if debug:
print(
f" Extracted '{letter}' using fallback 'last_valid_letter'"
)
return letter, "fallback_last_valid_letter"
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."""
# Format the prompt
prompt, choices, target_idx = format_bbh_prompt(item["raw_item"])
num_choices = len(choices)
gold_letter = (
ascii_uppercase[target_idx] if 0 <= target_idx < num_choices else None
)
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, num_choices, choices, debug=self.config.full_debug
)
# Score
is_correct = (
extracted_answer == gold_letter
if extracted_answer and gold_letter
else False
)
if self.config.full_debug:
print(f"\n--- Item: {item['id']} ---")
print(f"Subset: {item['subset']}")
print(f"Input: {item['input'][:100]}...")
print(
f"Gold: {gold_letter}, Extracted: {extracted_answer} (method: {extraction_method})"
)
print(f"Correct: {is_correct}")
return {
"item_id": item["id"],
"subset": item["subset"],
"input": item["input"][:200],
"num_choices": num_choices,
"gold_letter": gold_letter,
"extracted_answer": extracted_answer,
"extraction_method": extraction_method,
"is_correct": is_correct,
"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 BBH evaluation."""
print(f"\n{'='*60}")
print("Starting BBH 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 BBH")
# 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 overall 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-subset metrics
subset_metrics = {}
for r in valid_results:
subset = r.get("subset", "unknown")
if subset not in subset_metrics:
subset_metrics[subset] = {"total": 0, "correct": 0}
subset_metrics[subset]["total"] += 1
if r["is_correct"]:
subset_metrics[subset]["correct"] += 1
for subset in subset_metrics:
s_total = subset_metrics[subset]["total"]
s_correct = subset_metrics[subset]["correct"]
subset_metrics[subset]["accuracy"] = (
s_correct / s_total if s_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))
# Extraction method breakdown
method_counts = {}
for r in valid_results:
method = r.get("extraction_method", "unknown")
method_counts[method] = method_counts.get(method, 0) + 1
metrics = {
"accuracy": accuracy,
"total_evaluated": total,
"total_correct": correct,
"num_subsets": len(subset_metrics),
"format_compliance_rate": format_valid / total if total > 0 else 0.0,
"thinking_utilization_rate": has_thinking / total if total > 0 else 0.0,
"subset_metrics": subset_metrics,
"extraction_methods": method_counts,
}
print(f"\n{'='*60}")
print("BBH Evaluation Results")
print(f"{'='*60}")
print(f" Overall Accuracy: {accuracy:.2%} ({correct}/{total})")
print(f" Number of Subsets: {len(subset_metrics)}")
print(f" Format Compliance: {format_valid / total:.2%}")
if self.config.thinking_mode:
print(f" Thinking Utilization: {has_thinking / total:.2%}")
print("\n Per-Subset Breakdown:")
for subset, data in sorted(
subset_metrics.items(), key=lambda x: -x[1]["accuracy"]
):
print(
f" {subset}: {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 = {
"bbh/accuracy": metrics.get("accuracy", 0),
"bbh/total_evaluated": metrics.get("total_evaluated", 0),
"bbh/num_subsets": metrics.get("num_subsets", 0),
"bbh/format_compliance_rate": metrics.get("format_compliance_rate", 0),
"bbh/thinking_utilization_rate": metrics.get(
"thinking_utilization_rate", 0
),
}
# Log per-subset accuracies
for subset, data in metrics.get("subset_metrics", {}).items():
safe_name = subset.replace(" ", "_")[:40]
log_dict[f"bbh/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__":
BBHEvalEnv.cli()