plot_pass_at_k.py

import json
import numpy as np
import matplotlib.pyplot as plt

def load_graded(path):
    with open(path) as f:
        data = json.load(f)
    return [item["graded_list"] for item in data]

datasets = [
    ("Qwen3-4B-Base", "output/Qwen3-4B-Base/Scenario.codegeneration_16_0.6_eval_all.json"),
    ("ftajwar (MaxRL 1000)", "output/ftajwar/qwen3_4B_Base_MaxRL_Polaris_1000_steps/Scenario.codegeneration_16_0.6_eval_all.json"),
    ("ftajwar (GRPO 1000)", "output/ftajwar/qwen3_4B_Base_GRPO_Polaris_1000_steps/Scenario.codegeneration_16_0.6_eval_all.json"),
]

def bootstrap_pass_at_k(graded_lists, k, n_bootstrap=10000, rng=None):
    """Bootstrapping estimator for pass@k averaged over problems."""
    if rng is None:
        rng = np.random.default_rng(42)
    problem_scores = []
    for outcomes in graded_lists:
        outcomes_arr = np.array(outcomes, dtype=bool)
        n = len(outcomes_arr)
        # Sample k indices with replacement, check if any pass
        samples = rng.integers(0, n, size=(n_bootstrap, k))
        any_pass = outcomes_arr[samples].any(axis=1)
        problem_scores.append(any_pass.mean())
    return np.mean(problem_scores)

k_values = [1, 2, 4, 8, 16]
fig, ax = plt.subplots(figsize=(8, 5))

all_values = []
for label, path in datasets:
    graded = load_graded(path)
    rng = np.random.default_rng(42)
    pass_at_k = [bootstrap_pass_at_k(graded, k, rng=rng) for k in k_values]
    all_values.extend(pass_at_k)
    print(f"\n{label}:")
    for k, v in zip(k_values, pass_at_k):
        print(f"  pass@{k}: {v:.4f}")
    line, = ax.plot(k_values, pass_at_k, marker="o", linewidth=2, markersize=8, label=label)
    for k, v in zip(k_values, pass_at_k):
        ax.annotate(f"{v:.3f}", (k, v), textcoords="offset points", xytext=(5, 6),
                    fontsize=8, color=line.get_color())

ax.set_xscale("log", base=2)
ax.set_xticks(k_values)
ax.set_xticklabels([str(k) for k in k_values])
ax.set_xlabel("k", fontsize=12)
ax.set_ylabel("pass@k", fontsize=12)
ax.set_title("pass@k (bootstrapping estimator)", fontsize=13)
ax.set_ylim(0, max(all_values) * 1.3)
ax.legend(fontsize=10)
ax.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig("instruct.png", dpi=150)
print("\nSaved instruct.png")