scripts/build_poster.py

#!/usr/bin/env python3
"""Build poster_embed.html for TarGATE ICML logbook (matplotlib + data-URI)."""

from __future__ import annotations

import base64
import io
import json
from pathlib import Path

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyBboxPatch, Rectangle

ROOT = Path(__file__).resolve().parents[1]
OUT_DIR = ROOT / "outputs"
OUT_DIR.mkdir(parents=True, exist_ok=True)


def load_results() -> dict:
    p = OUT_DIR / "results.json"
    if p.exists():
        return json.loads(p.read_text())
    return {}


def card(ax, x, y, w, h, title, body, accent="#38bdf8"):
    ax.add_patch(
        FancyBboxPatch(
            (x, y),
            w,
            h,
            boxstyle="round,pad=0.02,rounding_size=0.08",
            facecolor="#1e293b",
            edgecolor="#475569",
            linewidth=1,
        )
    )
    ax.text(x + 0.15, y + h - 0.28, title, fontsize=11, color=accent, fontweight="bold", family="sans-serif")
    ax.text(x + 0.15, y + h / 2 - 0.15, body, fontsize=9.5, color="#e2e8f0", va="center", family="sans-serif")


def render_poster_png(results: dict) -> bytes:
    fig = plt.figure(figsize=(16, 9), dpi=140)
    ax = fig.add_axes([0, 0, 1, 1])
    ax.set_xlim(0, 16)
    ax.set_ylim(0, 9)
    ax.axis("off")

    ax.add_patch(Rectangle((0, 0), 16, 9, facecolor="#0f172a", edgecolor="none"))
    ax.add_patch(Rectangle((0, 7.7), 16, 1.3, facecolor="#1e3a5f", edgecolor="none"))
    ax.add_patch(Rectangle((0, 7.65), 16, 0.08, facecolor="#38bdf8", edgecolor="none"))

    ax.text(0.4, 8.55, "REPRODUCTION  ·  ICML 2026", fontsize=11, color="#7dd3fc", fontweight="bold", va="center")
    ax.text(0.4, 8.05, "TarGATE: Target-Aware Data Selection via Token-Attenuation Gates", fontsize=16, color="white", fontweight="bold", va="center")
    ax.text(15.6, 8.3, "OpenReview: xaqSrbGpPN\nICML poster #60688", fontsize=9, color="#cbd5e1", ha="right", va="center")

    sel = results.get("selection", {})
    tg = sel.get("TarGATE", {})
    rnd = sel.get("Random", {})
    nll = sel.get("NegNLL", {})
    sft = results.get("sft", {})
    xfer = results.get("transfer_sft", {})
    eff = results.get("efficiency", {})
    warm = results.get("warmup", {})

    def pct(m, k):
        v = m.get(k)
        return f"{100*v:.1f}%" if isinstance(v, (int, float)) else "—"

    outcome = (
        f"Outcome: TarGATE precision {pct(tg,'precision')} vs Random {pct(rnd,'precision')} / NegNLL {pct(nll,'precision')}  ·  "
        f"gates-only warmup  ·  T4 Job"
    )
    ax.add_patch(
        FancyBboxPatch((0.35, 6.85), 15.3, 0.65, boxstyle="round,pad=0.02,rounding_size=0.08", facecolor="#052e16", edgecolor="#4ade80", linewidth=1.5)
    )
    ax.text(8.0, 7.18, outcome, fontsize=10, color="#bbf7d0", ha="center", va="center", fontweight="bold")

    card(
        ax,
        0.35,
        4.7,
        7.5,
        1.95,
        "Claim 1 — Noisy selection (top-10%)",
        f"Precision (target in selected):\n"
        f"  TarGATE  {pct(tg,'precision')}   Random  {pct(rnd,'precision')}   NegNLL  {pct(nll,'precision')}\n"
        f"Recovery: TarGATE {pct(tg,'recovery')}  ·  n_selected={tg.get('n_selected','—')}\n"
        f"Pool: 800 GSM8K + 800 noise  ·  ref=80  ·  Qwen2.5-0.5B",
        "#38bdf8",
    )
    card(
        ax,
        8.15,
        4.7,
        7.5,
        1.95,
        "Claim 1 — Short SFT (optional signal)",
        f"Selector SFT eval acc (numeric match):\n"
        f"  TarGATE {sft.get('TarGATE',{}).get('eval_acc','—')}  "
        f"Random {sft.get('Random',{}).get('eval_acc','—')}  "
        f"NegNLL {sft.get('NegNLL',{}).get('eval_acc','—')}\n"
        f"Transfer 1.5B: TarGATE {xfer.get('TarGATE',{}).get('eval_acc','—')}  "
        f"Random {xfer.get('Random',{}).get('eval_acc','—')}\n"
        f"(Scaled SFT steps — mechanism verified primarily via precision)",
        "#a78bfa",
    )
    card(
        ax,
        0.35,
        2.5,
        7.5,
        1.95,
        "Claim 2 — Efficiency",
        f"Trainable gate params: {eff.get('trainable_gate_params', warm.get('n_trainable_gate_params','—'))}\n"
        f"Warmup wall: {warm.get('warmup_wall_sec', eff.get('warmup_time_sec','—'))}s\n"
        f"Score wall: TarGATE {eff.get('score_time_targate_sec','—')}s  vs  NLL {eff.get('score_time_nll_sec','—')}s\n"
        f"Base frozen; only linear IRR gates updated",
        "#fbbf24",
    )
    card(
        ax,
        8.15,
        2.5,
        7.5,
        1.95,
        "Claim 2 — Cross-model transfer",
        "Small selector (0.5B) curates data for larger (1.5B) SFT.\n"
        f"Selected precision (from small model): {pct(tg,'precision')}\n"
        "Data transfer (not weight transfer) matches paper claim that\n"
        "a smaller selector can curate fine-tuning data for larger LMs.",
        "#34d399",
    )

    ax.add_patch(FancyBboxPatch((0.35, 0.35), 15.3, 1.9, boxstyle="round,pad=0.02,rounding_size=0.08", facecolor="#1e293b", edgecolor="#475569", linewidth=1))
    ax.text(0.55, 1.9, "Method · Scope · Links", fontsize=11, color="#38bdf8", fontweight="bold")
    ax.text(
        0.55,
        1.1,
        "TarGATE = token-level IRR gates on FFN residual  ·  joint quality loss (↑IRR on reference, ↓IRR on noise)\n"
        "Scale: toy/medium vs paper (Tulu-200k / full LoRA). Job: HF t4-small  ·  Bucket: junwatu/targate-repro-artifacts\n"
        "Code: anonymous.4open.science/r/TarGATE-4008  ·  OpenReview: openreview.net/forum?id=xaqSrbGpPN\n"
        "Models: Qwen/Qwen2.5-0.5B-Instruct  ·  Qwen/Qwen2.5-1.5B-Instruct  ·  Data: openai/gsm8k",
        fontsize=9,
        color="#e2e8f0",
        va="center",
        family="sans-serif",
    )

    buf = io.BytesIO()
    fig.savefig(buf, format="png", facecolor=fig.get_facecolor(), bbox_inches="tight", pad_inches=0.05)
    plt.close(fig)
    return buf.getvalue()


def write_embed(png_bytes: bytes, path: Path) -> None:
    b64 = base64.b64encode(png_bytes).decode("ascii")
    # Hotspots map to claim page slugs
    html = f"""<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>Reproduction poster — TarGATE</title>
<style>
  .poster-wrap {{ position: relative; width: 100%; max-width: 1400px; margin: 0 auto; }}
  .poster-wrap img {{ width: 100%; height: auto; display: block; border-radius: 12px; }}
  .poster-hotspot {{
    position: absolute; border: 0; padding: 0; margin: 0;
    background: transparent; cursor: pointer; z-index: 2;
  }}
  .poster-hotspot:hover, .poster-hotspot:focus {{
    outline: 2px solid #38bdf8;
    background: rgba(56, 189, 248, 0.08) !important;
  }}
  .poster-hotspot-pill {{
    position: absolute; top: 8px; right: 8px;
    background: rgba(15, 23, 42, 0.85); color: #e2e8f0;
    font: 600 11px/1.2 system-ui, sans-serif;
    padding: 4px 8px; border-radius: 999px; opacity: 0.85;
    pointer-events: none;
  }}
  .poster-hotspot:hover .poster-hotspot-pill,
  .poster-hotspot:focus .poster-hotspot-pill {{ opacity: 1; }}
</style>
</head>
<body>
<div class="poster-wrap">
  <img src="data:image/png;base64,{b64}" alt="TarGATE reproduction poster"/>
  <button class="poster-hotspot" type="button"
    data-logbook-target="claim-1-targate-outperforms-related-baselines-in-both-noisy-and-real-world-data-selection-scenarios"
    style="left:2%;top:28%;width:47%;height:28%;"
    aria-label="Open Claim 1">
    <span class="poster-hotspot-pill">Open details ↗</span>
  </button>
  <button class="poster-hotspot" type="button"
    data-logbook-target="claim-1-targate-outperforms-related-baselines-in-both-noisy-and-real-world-data-selection-scenarios"
    style="left:51%;top:28%;width:47%;height:28%;"
    aria-label="Open Claim 1 SFT">
    <span class="poster-hotspot-pill">Open details ↗</span>
  </button>
  <button class="poster-hotspot" type="button"
    data-logbook-target="claim-2-the-method-exhibits-superior-computational-efficiency-and-strong-cross-model-transferability-for-curating-fine-tuning-data"
    style="left:2%;top:52%;width:47%;height:24%;"
    aria-label="Open Claim 2 efficiency">
    <span class="poster-hotspot-pill">Open details ↗</span>
  </button>
  <button class="poster-hotspot" type="button"
    data-logbook-target="claim-2-the-method-exhibits-superior-computational-efficiency-and-strong-cross-model-transferability-for-curating-fine-tuning-data"
    style="left:51%;top:52%;width:47%;height:24%;"
    aria-label="Open Claim 2 transfer">
    <span class="poster-hotspot-pill">Open details ↗</span>
  </button>
</div>
<script>
document.querySelectorAll('[data-logbook-target]').forEach(function(btn) {{
  btn.addEventListener('click', function() {{
    var slug = btn.getAttribute('data-logbook-target');
    if (window.parent && window.parent !== window) {{
      window.parent.postMessage({{ type: 'logbook-navigate', slug: slug }}, '*');
    }}
    location.hash = '#/' + slug;
  }});
}});
</script>
</body>
</html>
"""
    path.write_text(html)
    print(f"Wrote {path} ({path.stat().st_size} bytes)")


def main():
    results = load_results()
    png = render_poster_png(results)
    (OUT_DIR / "poster.png").write_bytes(png)
    write_embed(png, OUT_DIR / "poster_embed.html")
    # also to logbook dir if present
    lb = ROOT.parent / ".trackio" / "logbook"
    if lb.exists():
        write_embed(png, lb / "poster_embed.html")


if __name__ == "__main__":
    main()