scripts/make_plots.py

# /// script
# dependencies = ["plotly", "pandas"]
# ///
import json
import pandas as pd
import plotly.graph_objects as go

data = json.load(open('/Users/supreeth/xrpo/repro_xrpo/outputs/all_summaries.json'))

fig = go.Figure()
colors = {"grpo": "#4C78A8", "xrpo_full": "#F58518", "xrpo_no_icl": "#54A24B", "xrpo_no_novelty": "#B279A2"}
for name in ["grpo", "xrpo_full", "xrpo_no_icl", "xrpo_no_novelty"]:
    curve = data[name]["gsm8k_curve"]
    fig.add_trace(go.Scatter(
        x=[c["step"] for c in curve], y=[c["pass1"] for c in curve],
        mode="lines+markers", name=name, line=dict(color=colors[name]),
    ))
fig.update_layout(
    title="GSM8K pass@1 vs training step (Qwen3-0.6B, scaled repro)",
    xaxis_title="optimizer step", yaxis_title="pass@1",
    template="plotly_white", width=800, height=500,
)
fig.write_html("/Users/supreeth/xrpo/repro_xrpo/outputs/claim3_convergence.html")

rows = []
for name, d in data.items():
    for c in d["gsm8k_curve"]:
        rows.append({"config": name, **c})
pd.DataFrame(rows).to_csv("/Users/supreeth/xrpo/repro_xrpo/outputs/claim3_convergence.csv", index=False)

# final metrics table
rows2 = []
for name, d in data.items():
    g, m = d["final_gsm8k"], d["final_math500"]
    rows2.append({
        "config": name,
        "gsm8k_pass1": g["pass1"], "gsm8k_cons8": g["cons"], "gsm8k_avg_len": g["avg_len"],
        "math500_pass1": m["pass1"], "math500_cons8": m["cons"], "math500_avg_len": m["avg_len"],
        "avg_pass1": (g["pass1"]+m["pass1"])/2, "avg_cons8": (g["cons"]+m["cons"])/2,
        "avg_len": (g["avg_len"]+m["avg_len"])/2,
        "icl_bank_size": d.get("icl_bank_size"), "icl_injections_total": d.get("icl_injections_total"),
        "elapsed_min": d["elapsed_sec"]/60,
    })
df = pd.DataFrame(rows2)
df.to_csv("/Users/supreeth/xrpo/repro_xrpo/outputs/final_metrics.csv", index=False)
print(df.to_string(index=False))

# bar chart final pass@1 comparison
fig2 = go.Figure()
fig2.add_trace(go.Bar(x=df["config"], y=df["avg_pass1"], marker_color=[colors[c] for c in df["config"]], name="avg pass@1"))
fig2.update_layout(title="Final avg pass@1 (GSM8K+MATH-500) by config", template="plotly_white", width=700, height=450, yaxis_title="pass@1")
fig2.write_html("/Users/supreeth/xrpo/repro_xrpo/outputs/final_pass1_bar.html")
print("done")