File size: 9,253 Bytes
8d18b7c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 | """Comparative Evaluation Between Model Sizes and Architectures"""
import json
import logging
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
import torch
from scipy import stats
logger = logging.getLogger(__name__)
@dataclass
class ModelComparison:
"""Results of comparing two models."""
model_a_name: str
model_b_name: str
metrics_a: Dict[str, float]
metrics_b: Dict[str, float]
differences: Dict[str, float] = field(default_factory=dict)
statistical_tests: Dict[str, Dict[str, float]] = field(default_factory=dict)
summary: str = ""
def compute_differences(self):
"""Compute absolute and relative differences."""
self.differences = {}
for key in self.metrics_a:
if key in self.metrics_b:
diff = self.metrics_a[key] - self.metrics_b[key]
rel_diff = diff / (self.metrics_b[key] if self.metrics_b[key] != 0 else 1e-8)
self.differences[key] = {
"absolute": diff,
"relative": rel_diff,
}
def compute_statistical_tests(self, samples_a: List[float], samples_b: List[float]):
"""Run statistical significance tests."""
if len(samples_a) < 2 or len(samples_b) < 2:
return
# T-test
t_stat, t_p = stats.ttest_ind(samples_a, samples_b, equal_var=False)
# Mann-Whitney U test (non-parametric)
u_stat, u_p = stats.mannwhitneyu(samples_a, samples_b, alternative="two-sided")
self.statistical_tests = {
"t_test": {"statistic": float(t_stat), "p_value": float(t_p)},
"mann_whitney_u": {"statistic": float(u_stat), "p_value": float(u_p)},
}
def generate_summary(self) -> str:
"""Generate human-readable summary."""
lines = [
f"Model Comparison: {self.model_a_name} vs {self.model_b_name}",
"=" * 60,
]
for metric, diffs in self.differences.items():
a_val = self.metrics_a[metric]
b_val = self.metrics_b[metric]
abs_diff = diffs["absolute"]
rel_diff = diffs["relative"] * 100
if abs_diff > 0:
better = self.model_a_name if a_val > b_val else self.model_b_name
else:
better = "tie"
lines.append(f"{metric:30s}: {a_val:.4f} vs {b_val:.4f} "
f"(diff: {abs_diff:+.4f}, {rel_diff:+.1f}%) -> {better}")
lines.append("\nStatistical Significance:")
for test_name, results in self.statistical_tests.items():
p_val = results["p_value"]
sig = "significant" if p_val < 0.05 else "not significant"
lines.append(f" {test_name}: p={p_val:.4f} ({sig})")
self.summary = "\n".join(lines)
return self.summary
class ComparativeEvaluator:
"""Evaluate and compare multiple models."""
def __init__(
self,
models: Dict[str, torch.nn.Module],
tokenizers: Dict[str, Any],
benchmark_config: Any,
):
self.models = models
self.tokenizers = tokenizers
self.config = benchmark_config
def compare_models(
self,
model_names: List[str],
benchmark_datasets: Dict[str, Any],
) -> ModelComparison:
"""Compare two models on multiple benchmarks."""
if len(model_names) != 2:
raise ValueError("Can only compare exactly 2 models")
name_a, name_b = model_names
model_a = self.models[name_a]
model_b = self.models[name_b]
tokenizer_a = self.tokenizers[name_a]
tokenizer_b = self.tokenizers[name_b]
# Run benchmarks
metrics_a = self._run_benchmarks(model_a, tokenizer_a, benchmark_datasets)
metrics_b = self._run_benchmarks(model_b, tokenizer_b, benchmark_datasets)
comparison = ModelComparison(
model_a_name=name_a,
model_b_name=name_b,
metrics_a=metrics_a,
metrics_b=metrics_b,
)
comparison.compute_differences()
# Note: statistical tests would require multiple runs/samples
comparison.generate_summary()
return comparison
def _run_benchmarks(
self,
model: torch.nn.Module,
tokenizer: Any,
datasets: Dict[str, Any],
) -> Dict[str, float]:
"""Run all benchmarks on a model."""
from .benchmark import BenchmarkSuite, BenchmarkConfig
config = BenchmarkConfig(
batch_size=self.config.batch_size,
max_seq_length=self.config.max_seq_length,
datasets=list(datasets.keys()),
)
suite = BenchmarkSuite(model, tokenizer, config)
results = suite.run_all_benchmarks()
# Flatten results
flat_metrics = {}
for category, metrics in results["benchmarks"].items():
if isinstance(metrics, dict):
for key, value in metrics.items():
if isinstance(value, (int, float)):
flat_metrics[f"{category}_{key}"] = value
elif isinstance(metrics, (int, float)):
flat_metrics[category] = metrics
return flat_metrics
def generate_comparison_report(
self,
comparisons: List[ModelComparison],
output_path: str,
):
"""Generate comprehensive comparison report."""
report = {
"timestamp": torch.datetime.now().isoformat(),
"comparisons": [],
}
for comp in comparisons:
report["comparisons"].append({
"models": [comp.model_a_name, comp.model_b_name],
"metrics_a": comp.metrics_a,
"metrics_b": comp.metrics_b,
"differences": comp.differences,
"statistical_tests": comp.statistical_tests,
"summary": comp.summary,
})
with open(output_path, 'w') as f:
json.dump(report, f, indent=2)
logger.info(f"Comparison report saved to {output_path}")
def compare_model_sizes(
models: Dict[str, torch.nn.Module],
tokenizers: Dict[str, Any],
config: Any,
output_dir: str,
) -> Dict[str, ModelComparison]:
"""Compare 7B vs 32B vs 70B models."""
comparisons = {}
evaluator = ComparativeEvaluator(models, tokenizers, config)
# Load benchmark datasets
from .eval_datasets import load_human_eval, load_gsm8k, load_truthfulqa
datasets = {
"human_eval": load_human_eval()[:100],
"gsm8k": load_gsm8k()[:100],
"truthfulqa": load_truthfulqa()[:100],
}
# Compare all pairs
model_names = list(models.keys())
for i in range(len(model_names)):
for j in range(i + 1, len(model_names)):
pair = (model_names[i], model_names[j])
logger.info(f"Comparing {pair[0]} vs {pair[1]}")
comparison = evaluator.compare_models(list(pair), datasets)
comparisons[f"{pair[0]}_vs_{pair[1]}"] = comparison
# Save individual comparison
output_path = f"{output_dir}/comparison_{pair[0]}_vs_{pair[1]}.json"
evaluator.generate_comparison_report([comparison], output_path)
return comparisons
def analyze_scaling_laws(comparisons: Dict[str, ModelComparison]) -> Dict[str, Any]:
"""Analyze scaling laws from model comparisons."""
# Extract size vs performance data
sizes = [] # In parameters (B)
perplexities = []
accuracies = []
code_scores = []
# Map model names to sizes (this would come from configs)
size_map = {"zenith-7b": 7, "zenith-32b": 32, "zenith-70b": 70}
for comp_key, comp in comparisons.items():
# For each comparison, extract metrics
for metric, value in comp.metrics_a.items():
if "perplexity" in metric:
model_name = comp.model_a_name
if model_name in size_map:
sizes.append(size_map[model_name])
perplexities.append(value)
elif "accuracy" in metric or "pass@1" in metric:
model_name = comp.model_a_name
if model_name in size_map:
accuracies.append((size_map[model_name], value))
# Compute scaling exponents (power law fit)
if len(sizes) >= 2 and len(perplexities) >= 2:
log_sizes = np.log(sizes)
log_ppl = np.log(perplexities)
slope, intercept, r_value, p_value, std_err = stats.linregress(log_sizes, log_ppl)
scaling_exponent = -slope # Negative because larger models have lower perplexity
else:
scaling_exponent = None
return {
"sizes": sizes,
"perplexities": perplexities,
"accuracies": accuracies,
"scaling_exponent": scaling_exponent,
"r_squared": r_value**2 if scaling_exponent is not None else None,
}
|