GitHub Action
Clean deployment without binary files
f884e6e
#!/usr/bin/env python3
"""
Deterministic Evaluation Controls
Provides utilities for ensuring reproducible and deterministic evaluation results
across different runs and environments.
"""
import logging
import os
import random
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional
logger = logging.getLogger(__name__)
# Default seed for reproducible evaluations
DEFAULT_EVALUATION_SEED = 42
@dataclass
class DeterministicConfig:
"""Configuration for deterministic evaluation settings."""
# Random seed for reproducibility
random_seed: int = DEFAULT_EVALUATION_SEED
# Sort results for consistent ordering
sort_results: bool = True
# Use fixed precision for floating point comparisons
float_precision: int = 6
# Consistent evaluation order
consistent_order: bool = True
# Additional deterministic flags
deterministic_mode: bool = True
# Environment variables to set for reproducibility
env_vars: Dict[str, str] = field(
default_factory=lambda: {
"PYTHONHASHSEED": "0",
"CUBLAS_WORKSPACE_CONFIG": ":4096:8",
}
)
class DeterministicEvaluator:
"""
Wrapper that ensures deterministic evaluation behavior.
Provides:
- Fixed random seeds
- Consistent result ordering
- Reproducible floating point precision
- Environment variable controls
"""
def __init__(self, config: Optional[DeterministicConfig] = None):
"""Initialize deterministic evaluator with configuration."""
self.config = config or DeterministicConfig()
self._setup_deterministic_environment()
def _setup_deterministic_environment(self) -> None:
"""Configure environment for deterministic behavior."""
# Set random seed
random.seed(self.config.random_seed)
# Set environment variables for reproducibility
for key, value in self.config.env_vars.items():
os.environ[key] = value
# Try to set numpy seed if available
try:
import numpy as np
np.random.seed(self.config.random_seed)
except ImportError:
logger.warning("numpy is not installed; deterministic seeding for numpy will be skipped.")
# Try to set torch seed if available
try:
import torch
torch.manual_seed(self.config.random_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(self.config.random_seed)
torch.cuda.manual_seed_all(self.config.random_seed)
# Enable deterministic algorithms in PyTorch
torch.use_deterministic_algorithms(True, warn_only=True)
except ImportError:
# torch is optional; skip deterministic seeding if not installed
pass
logger.info(f"Deterministic environment configured with seed: {self.config.random_seed}")
def normalize_float(self, value: float) -> float:
"""Normalize floating point value to consistent precision."""
return round(value, self.config.float_precision)
def normalize_metrics(self, metrics: Dict[str, Any]) -> Dict[str, Any]:
"""Normalize metric values for consistent precision."""
normalized = {}
for key, value in metrics.items():
if isinstance(value, float):
normalized[key] = self.normalize_float(value)
elif isinstance(value, dict):
normalized[key] = self.normalize_metrics(value)
else:
normalized[key] = value
return normalized
def sort_evaluation_results(
self, results: List[Dict[str, Any]], sort_key: str = "query_id"
) -> List[Dict[str, Any]]:
"""Sort evaluation results for consistent ordering."""
if not self.config.sort_results:
return results
try:
return sorted(results, key=lambda x: x.get(sort_key, ""))
except (KeyError, TypeError):
# Fallback to string representation if sort key issues
return sorted(results, key=str)
def ensure_deterministic_order(self, items: List[Any], key_func=None) -> List[Any]:
"""Ensure consistent ordering of items."""
if not self.config.consistent_order:
return items
if key_func:
return sorted(items, key=key_func)
# Try natural sorting, fall back to string representation
try:
return sorted(items)
except TypeError:
return sorted(items, key=str)
def create_deterministic_groundedness_evaluator(seed: Optional[int] = None) -> DeterministicEvaluator:
"""Create a deterministic evaluator specifically configured for groundedness evaluation."""
config = DeterministicConfig(
random_seed=seed or DEFAULT_EVALUATION_SEED,
sort_results=True,
float_precision=4, # Slightly lower precision for groundedness scores
consistent_order=True,
deterministic_mode=True,
)
return DeterministicEvaluator(config)
def evaluate_groundedness_deterministic(
generated_text: str, source_passages: List[str], evaluator: Optional[DeterministicEvaluator] = None
) -> Dict[str, float]:
"""
Evaluate groundedness with deterministic behavior.
Uses token overlap and passage-level matching with consistent ordering
and normalized precision.
"""
if evaluator is None:
evaluator = create_deterministic_groundedness_evaluator()
if not generated_text.strip() or not source_passages:
return evaluator.normalize_metrics(
{"groundedness_score": 0.0, "passage_coverage": 0.0, "token_overlap": 0.0, "exact_matches": 0.0}
)
# Normalize inputs for consistent processing
generated_tokens = set(generated_text.lower().split())
# Process passages in consistent order
sorted_passages = evaluator.ensure_deterministic_order(source_passages)
# Calculate passage-level scores
passage_scores = []
total_coverage = 0
exact_matches = 0
for passage in sorted_passages:
if not passage.strip():
continue
passage_tokens = set(passage.lower().split())
# Token overlap for this passage
if passage_tokens:
overlap = len(generated_tokens & passage_tokens) / len(passage_tokens)
passage_scores.append(overlap)
# Check for exact phrase matches (deterministic substring matching)
passage_lower = passage.lower()
generated_lower = generated_text.lower()
# Count exact matches using consistent methodology
exact_phrases = []
words = generated_lower.split()
for i in range(len(words)):
for j in range(i + 2, min(i + 8, len(words) + 1)): # 2-7 word phrases
phrase = " ".join(words[i:j])
if phrase in passage_lower and phrase not in exact_phrases:
exact_phrases.append(phrase)
if exact_phrases:
exact_matches += 1
total_coverage += overlap
# Calculate aggregate scores with normalization
if passage_scores:
groundedness_score = sum(passage_scores) / len(passage_scores)
passage_coverage = total_coverage / len(sorted_passages)
else:
groundedness_score = 0.0
passage_coverage = 0.0
# Overall token overlap across all passages
all_source_tokens = set()
for passage in sorted_passages:
all_source_tokens.update(passage.lower().split())
if all_source_tokens:
token_overlap = len(generated_tokens & all_source_tokens) / len(all_source_tokens)
else:
token_overlap = 0.0
exact_match_rate = exact_matches / len(sorted_passages) if sorted_passages else 0.0
metrics = {
"groundedness_score": groundedness_score,
"passage_coverage": passage_coverage,
"token_overlap": token_overlap,
"exact_matches": exact_match_rate,
}
return evaluator.normalize_metrics(metrics)
def evaluate_citation_accuracy_deterministic(
generated_text: str,
returned_sources: List[Dict[str, Any]],
expected_sources: List[str],
evaluator: Optional[DeterministicEvaluator] = None,
) -> Dict[str, float]:
"""
Evaluate citation accuracy with deterministic behavior.
Provides consistent filename matching and source validation.
"""
if evaluator is None:
evaluator = create_deterministic_groundedness_evaluator()
if not expected_sources:
# If no expected sources, score based on whether any sources were returned
return evaluator.normalize_metrics(
{
"citation_accuracy": 1.0 if not returned_sources else 0.0,
"source_precision": 1.0,
"source_recall": 1.0,
"exact_filename_matches": 1.0,
}
)
# Normalize filenames for consistent matching
def normalize_filename(filename: str) -> str:
"""Normalize filename for consistent comparison."""
if not filename:
return ""
import os
import re
# Remove query parameters and fragments
filename = re.sub(r"[?#].*$", "", filename.strip())
# Get basename
basename = os.path.basename(filename)
# Remove common extensions consistently
basename = re.sub(
r"\.(md|markdown|txt|html|htm|pdf|csv|json|yaml|yml|py|ipynb)$", "", basename, flags=re.IGNORECASE
)
return basename.lower()
# Extract returned filenames in consistent order
returned_filenames = set()
sorted_sources = evaluator.ensure_deterministic_order(returned_sources, key_func=str)
for source in sorted_sources:
if isinstance(source, dict):
candidates = [source.get(k) for k in ["filename", "source_file", "file", "url", "path", "source"]]
# Check metadata
metadata = source.get("metadata", {})
if isinstance(metadata, dict):
candidates.extend([metadata.get(k) for k in ["filename", "file", "source_file"]])
else:
candidates = [str(source)]
for candidate in candidates:
if candidate:
normalized = normalize_filename(str(candidate))
if normalized:
returned_filenames.add(normalized)
# Normalize expected sources
expected_normalized = set()
sorted_expected = evaluator.ensure_deterministic_order(expected_sources)
for expected in sorted_expected:
normalized = normalize_filename(str(expected))
if normalized:
expected_normalized.add(normalized)
# Calculate matches with consistent methodology
exact_matches = len(expected_normalized & returned_filenames)
# Calculate precision and recall
if returned_filenames:
precision = exact_matches / len(returned_filenames)
else:
precision = 1.0 if not expected_normalized else 0.0
if expected_normalized:
recall = exact_matches / len(expected_normalized)
else:
recall = 1.0
# Overall citation accuracy (F1-like score)
if precision + recall > 0:
citation_accuracy = 2 * (precision * recall) / (precision + recall)
else:
citation_accuracy = 0.0
exact_filename_match_rate = recall # Same as recall for exact matches
metrics = {
"citation_accuracy": citation_accuracy,
"source_precision": precision,
"source_recall": recall,
"exact_filename_matches": exact_filename_match_rate,
}
return evaluator.normalize_metrics(metrics)
# Utility functions for integration
def setup_deterministic_evaluation(seed: Optional[int] = None) -> DeterministicEvaluator:
"""Setup deterministic evaluation environment."""
return create_deterministic_groundedness_evaluator(seed)
def get_evaluation_seed() -> int:
"""Get the evaluation seed from environment or use default."""
try:
return int(os.getenv("EVALUATION_SEED", DEFAULT_EVALUATION_SEED))
except (ValueError, TypeError):
return DEFAULT_EVALUATION_SEED