File size: 4,923 Bytes
6cf60f3 | 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 | 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]
def bootstrap_pass_at_k(graded_lists, k, n_bootstrap=10000, rng=None):
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)
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)
with open("output/base_model_evals.json") as f:
base_evals = json.load(f)
one_shot = base_evals["results"]["generic_base_one_shot"]
maxrl_steps = [100, 200, 300, 400]
tailrl_steps = [100, 200, 300, 400]
k_values = [1, 2, 4, 8, 16]
left_k_values = [1, 16]
maxrl_data = {}
for step in maxrl_steps:
path = f"output/maxrl_binary_{step}/Scenario.codegeneration_16_0.6_eval_all.json"
graded = load_graded(path)
rng = np.random.default_rng(42)
maxrl_data[step] = {k: bootstrap_pass_at_k(graded, k, rng=rng) for k in k_values}
tailrl_data = {}
for step in tailrl_steps:
path = f"output/tailrl_cont_{step}/Scenario.codegeneration_16_0.6_eval_all.json"
graded = load_graded(path)
rng = np.random.default_rng(42)
tailrl_data[step] = {k: bootstrap_pass_at_k(graded, k, rng=rng) for k in k_values}
results_json = {
"base_one_shot": {f"pass@{k}": one_shot[f"pass@{k}"] for k in k_values},
"maxrl_binary": {str(step): {f"pass@{k}": float(v) for k, v in ks.items()} for step, ks in maxrl_data.items()},
"tailrl_cont": {str(step): {f"pass@{k}": float(v) for k, v in ks.items()} for step, ks in tailrl_data.items()},
}
results_path = "output/pass_at_k_results.json"
with open(results_path, "w") as f:
json.dump(results_json, f, indent=2)
print(f"Saved {results_path}")
fig, (ax_left, ax_right) = plt.subplots(1, 2, figsize=(14, 5))
maxrl_colors = {'p@1': 'steelblue', 'p@16': 'navy'}
tailrl_colors = {'p@1': 'darkorange', 'p@16': 'saddlebrown'}
markers = {1: 'o', 16: 's'}
# --- Left plot: p@1 and p@16 vs steps ---
for k in left_k_values:
vals_maxrl = [maxrl_data[s][k] for s in maxrl_steps]
vals_tailrl = [tailrl_data[s][k] for s in tailrl_steps]
ax_left.plot(maxrl_steps, vals_maxrl, marker=markers[k], color=maxrl_colors[f'p@{k}'],
linewidth=2, markersize=7, label=f"maxrl p@{k}")
ax_left.plot(tailrl_steps, vals_tailrl, marker=markers[k], color=tailrl_colors[f'p@{k}'],
linewidth=2, markersize=7, linestyle='--', label=f"tailrl p@{k}")
for k, ls in zip(left_k_values, ['-', '--']):
ax_left.axhline(one_shot[f"pass@{k}"], color='gray', linewidth=1.5,
linestyle=ls, label=f"base one-shot p@{k}")
ax_left.set_xlabel("Training Steps", fontsize=12)
ax_left.set_ylabel("pass@k", fontsize=12)
ax_left.set_title("Performance vs Training Steps", fontsize=13)
ax_left.set_xticks(sorted(set(maxrl_steps + tailrl_steps)))
ax_left.legend(fontsize=10, loc="lower right")
ax_left.grid(True, alpha=0.3)
# --- Right plot: pass@k at last checkpoint (step 400) ---
last_maxrl = [maxrl_data[400][k] for k in k_values]
last_tailrl = [tailrl_data[400][k] for k in k_values]
base_one_shot_vals = [one_shot[f"pass@{k}"] for k in k_values]
ax_right.plot(k_values, base_one_shot_vals, marker='^', color='gray',
linewidth=2, markersize=8, linestyle='--', label="base one-shot")
for k, v in zip(k_values, base_one_shot_vals):
ax_right.annotate(f"{v:.3f}", (k, v), textcoords="offset points", xytext=(5, 6),
fontsize=8, color='gray')
ax_right.plot(k_values, last_maxrl, marker='o', color='steelblue',
linewidth=2, markersize=8, label="maxrl_binary_400")
ax_right.plot(k_values, last_tailrl, marker='s', color='darkorange',
linewidth=2, markersize=8, label="tailrl_cont_400")
for k, v in zip(k_values, last_maxrl):
ax_right.annotate(f"{v:.3f}", (k, v), textcoords="offset points", xytext=(5, 6),
fontsize=8, color='steelblue')
for k, v in zip(k_values, last_tailrl):
ax_right.annotate(f"{v:.3f}", (k, v), textcoords="offset points", xytext=(5, -14),
fontsize=8, color='darkorange')
ax_right.set_xscale("log", base=2)
ax_right.set_xticks(k_values)
ax_right.set_xticklabels([str(k) for k in k_values])
ax_right.set_xlabel("k", fontsize=12)
ax_right.set_ylabel("pass@k", fontsize=12)
ax_right.set_title("pass@k at Last Checkpoint", fontsize=13)
all_right = last_maxrl + last_tailrl + base_one_shot_vals
ax_right.set_ylim(min(all_right) * 0.9, max(all_right) * 1.15)
ax_right.legend(fontsize=10)
ax_right.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig("output/pass_at_k_comparison.png", dpi=150)
print("Saved output/pass_at_k_comparison.png")
|