make_heatmap.py

"""Render the correlation heatmap (FinePhrase Figure 22 style).

Columns are the 20 downstream metrics grouped into 7 families. Rows are four
predictor blocks: DCLM, Edu, and embedding-based diversity (the playbook's own
predictors, for reference), then the G-Vendi block (Published, Reproduced, Base
model, and Base model scored within-prompt). Every row but the last is the
raw-pooled Spearman; the last is the within-prompt z-scored version. Cell colour
is the correlation, the star is significance. All rows are computed on the same
83-cell grid (joined by run), so they are directly comparable.

Inputs (relative to the artifact root):
  data/pub.json, data/{repro,prx06bb}/*.json
  data/reference_predictors.json   (per-run DCLM / Edu / diversity values)
  data/downstream_scores.json      (per-run downstream scores, Figure 22 source)

Writes correlation_heatmap.png. Needs matplotlib + numpy + scipy.
Run: python make_heatmap.py
"""
import glob, json, math
from collections import defaultdict

import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyBboxPatch
from scipy.stats import spearmanr

# ---------------------------------------------------------------- column layout
COLUMN_GROUPS = [
    ("OVERALL", [("macro", "Macro Avg", True), ("micro", "Micro Avg", True)]),
    ("KNOWLEDGE", [
        ("GK", "GK Agg", True), ("arc", "ARC Easy", False), ("mmlu", "MMLU Redux", False)]),
    ("READING", [
        ("RC", "RC Agg", True), ("squad", "SQuAD v2", False), ("drop", "DROP", False)]),
    ("REASONING", [
        ("RES", "RES Agg", True), ("obqa", "OpenBookQA", False), ("xcsqa", "XCSQA", False)]),
    ("NLU", [
        ("NLU", "NLU Agg", True), ("wino", "WinoGrande", False),
        ("piqa", "PIQA", False), ("hella", "HellaSwag", False)]),
    ("MATH", [("MATH", "Math Agg", True), ("gsm8k", "GSM8K", False)]),
    ("TABLE", [
        ("TABLE", "Table Agg", True), ("wikitab", "WikiTableQ", False), ("treb", "TriviaQA", False)]),
]

SHORT_TO_COL = {
    "macro": "agg_score_macro", "micro": "agg_score_micro", "RC": "agg_score_RC",
    "GK": "agg_score_GK", "NLU": "agg_score_NLU", "MATH": "agg_score_MATH",
    "TABLE": "agg_score_TABLE", "RES": "agg_score_RES",
    "arc": "lighteval|arc_cf:easy|3/prob_norm_token",
    "drop": "lighteval|drop|3/prob_norm_token",
    "gsm8k": "lighteval|gsm8k|3/prob_norm_token",
    "hella": "lighteval|hellaswag_cf|3/prob_norm_token",
    "obqa": "lighteval|openbookqa_cf|3/prob_norm_token",
    "piqa": "lighteval|piqa_cf|3/prob_norm_token",
    "squad": "lighteval|squad_v2|3/prob_norm_token",
    "treb": "lighteval|treb_qa|3/prob_norm_token",
    "wikitab": "lighteval|wikitablequestions|3/prob_norm_token",
    "wino": "lighteval|winogrande_cf|3/prob_norm_token",
    "xcsqa": "lighteval|xcsqa_cf|3/prob_norm_token",
    "mmlu": "lighteval|mmlu_redux_cf:_average|3/prob_norm_token",
}

EXCLUDED_FOOTNOTE1 = {
    "format/article:1b-hq", "format/commentary:1b-hq", "format/discussion:1b-hq",
    "format/tutorial:1b-hq", "format/tutorial:12b-hq", "format/faq:1b-lq", "format/faq:12b-lq",
}

# ---------------------------------------------------------------- load the grid
# Downstream scores from the playbook's Figure 22 source (rephrasing_metadata.json),
# keyed by run = "bucket/prompt-suffix". Stored under CSV-style column names, so we
# translate to our short names via SHORT_TO_COL.
_downstream = json.load(open("data/downstream_scores.json"))

def lookup_scores(bucket, prompt, suffix):
    run = f"{bucket}/{prompt}-{suffix}"
    if run in _downstream:
        return {short: _downstream[run][col] for short, col in SHORT_TO_COL.items()}
    return None

def load_by_cell(globpat):
    out = {}
    for fp in glob.glob(globpat):
        d = json.load(open(fp))
        out[d["cell"]] = float(d["g_vendi"])
    return out

pub_gv = json.load(open("data/pub.json"))
ref = json.load(open("data/reference_predictors.json"))
gv = {
    "repro": load_by_cell("data/repro/*.json"),
    "prx06bb": load_by_cell("data/prx06bb/*.json"),
}

rows = []  # (bucket, prompt, {row_id: value}, {short: score})
for cell_full in sorted(gv["repro"]):
    if cell_full in EXCLUDED_FOOTNOTE1:
        continue
    bucket_prompt, suffix = cell_full.rsplit(":", 1)
    bucket, prompt = bucket_prompt.split("/", 1)
    pub_key = f"{bucket}/{prompt}-{suffix}"
    if pub_key not in pub_gv or pub_key not in ref:
        continue
    scores = lookup_scores(bucket, prompt, suffix)
    if scores is None:
        continue
    vals = {"pub": float(pub_gv[pub_key]["g_vendi_score"])}
    vals.update({p: gv[p][cell_full] for p in gv})
    vals.update({k: (float(v) if v is not None else float("nan")) for k, v in ref[pub_key].items()})
    rows.append((bucket, prompt, vals, scores))

print(f"matched cells: {len(rows)}")

# ---------------------------------------------------------------- correlations
def raw_corr(row_id, short):
    g = np.array([r[2][row_id] for r in rows], float)
    m = np.array([r[3][short] for r in rows], float)
    ok = ~(np.isnan(g) | np.isnan(m))
    res = spearmanr(g[ok], m[ok])
    return float(res.statistic), float(res.pvalue)

def zwithin_corr(row_id, short):
    byp = defaultdict(list)
    for r in rows:
        a, b = r[2][row_id], r[3][short]
        if not (math.isnan(a) or math.isnan(b)):
            byp[f"{r[0]}/{r[1]}"].append((a, b))
    zg, zm = [], []
    for v in byp.values():
        n = len(v)
        if n < 3:
            continue
        gs = [x[0] for x in v]; ms = [x[1] for x in v]
        mg, mm = sum(gs) / n, sum(ms) / n
        sg = math.sqrt(sum((x - mg) ** 2 for x in gs) / n)
        sm = math.sqrt(sum((x - mm) ** 2 for x in ms) / n)
        for a, b in v:
            zg.append((a - mg) / sg if sg else 0.0)
            zm.append((b - mm) / sm if sm else 0.0)
    res = spearmanr(zg, zm)
    return float(res.statistic), float(res.pvalue)

def stars(p):
    if p < 0.001: return "***"
    if p < 0.01:  return "**"
    if p < 0.05:  return "*"
    return ""


# ---------------------------------------------------------------- draw (mirrors the playbook's D3 embed)
flat_cols = [(s, lbl, agg) for _, members in COLUMN_GROUPS for (s, lbl, agg) in members]
N_COL = len(flat_cols)
CELL_W, CELL_H = 1.0, 0.82
ROUND = 0.085
GROUP_GAP = 0.34       # gap between row-groups

def col_x(i):
    return i * CELL_W

col_group_spans, _c = [], 0
for _t, _members in COLUMN_GROUPS:
    col_group_spans.append((_t, _c, len(_members)))
    _c += len(_members)
total_w = N_COL * CELL_W

# colours: d3.scaleDiverging([-0.85,0,0.85], interpolateRdBu) on cellColor(-r),
# then alpha-faded toward the white page near rho=0.
RdBu = plt.get_cmap("RdBu")
def cell_rgb(r):
    cval = float(np.clip((0.85 - r) / 1.7, 0.0, 1.0))
    rgb = np.array(RdBu(cval)[:3])
    alpha = float(np.clip(abs(r) / 0.85 * 1.8, 0.12, 1.0))
    return tuple(rgb * alpha + (1.0 - alpha))

DARK, MUTED, DIV = "#4d4d4d", "#a6a6a6", "#8c8c8c"


# Rows for the figure: reference predictors (raw) for context, then the proxy
# comparison: Published / Reproduced / Base model on raw, and Base model again
# on the within-prompt metric. corr is "raw" or "zw" per row.
CORR = {"raw": raw_corr, "zw": zwithin_corr}
FIG_ROWS = [
    ("DCLM", [("output_dclm_score", "Output DCLM", "raw"), ("input_dclm_score", "Input DCLM", "raw"),
              ("dclm_score_difference", "DCLM Δ", "raw"), ("dclm_score_improvement", "DCLM Improv %", "raw")]),
    ("Edu", [("output_edu_score", "Output Edu", "raw"), ("input_edu_score", "Input Edu", "raw"),
             ("edu_score_difference", "Edu Δ", "raw"), ("edu_score_improvement", "Edu Improv %", "raw")]),
    ("Diversity", [("vendi_score", "Vendi Score", "raw"), ("mean_intra_cos_sim", "Mean cos sim", "raw"),
                   ("near_dup_rate", "Near-dup rate", "raw")]),
    ("G-Vendi", [("pub", "Published", "raw"), ("repro", "Reproduced", "raw"),
                 ("prx06bb", "Base model", "raw"), ("prx06bb", "Base (within-prompt)", "zw")]),
]


def render(row_groups, outfile, title):
    """One heatmap: predictor rows (each with its own metric) vs the 20 benchmarks."""
    placed, spans, y = [], [], 0.0
    for gi, (gname, members) in enumerate(row_groups):
        gstart = y
        for row_id, label, metric in members:
            placed.append((row_id, label, metric, y))
            y += CELL_H
        spans.append((gname, gstart, y))
        if gi < len(row_groups) - 1:
            y += GROUP_GAP
    total_h = y

    width_span = total_w + 2.9
    height_span = total_h + 6.8
    SCALE = 0.42
    fig, ax = plt.subplots(figsize=(width_span * SCALE, height_span * SCALE))

    for row_id, plabel, metric, yc in placed:
        for ci, (short, label, agg) in enumerate(flat_cols):
            rho, p = CORR[metric](row_id, short)
            xc = col_x(ci)
            ax.add_patch(FancyBboxPatch(
                (xc + 0.06, yc + 0.06), CELL_W - 0.12, CELL_H - 0.12,
                boxstyle=f"round,pad=0,rounding_size={ROUND}",
                facecolor=cell_rgb(rho), edgecolor="white", linewidth=0.8))
            tcol = "white" if abs(rho) > 0.45 else DARK
            ax.text(xc + CELL_W / 2, yc + CELL_H / 2, f"{rho:.2f}", ha="center",
                    va="center", fontsize=8.3, color=tcol,
                    fontweight="bold" if abs(rho) > 0.4 else "normal")
            st = stars(p)
            if st:
                ax.text(xc + CELL_W - 0.1, yc + 0.16, st, ha="right", va="center",
                        fontsize=6.5, fontweight="bold",
                        color=("white" if abs(rho) > 0.45 else MUTED))

    # vertical dividers between column groups
    for gi, (t, start, n) in enumerate(col_group_spans):
        if gi > 0:
            x = col_x(start)
            ax.plot([x, x], [-0.04, total_h + 0.02], color=DIV,
                    linewidth=1.5 if gi == 1 else 0.9,
                    linestyle="-" if gi == 1 else (0, (4, 3)), zorder=0)

    # horizontal dividers between row-groups
    for gi, (gname, gs, ge) in enumerate(spans):
        if gi > 0:
            yd = gs - GROUP_GAP / 2
            ax.plot([-0.02, total_w + 0.02], [yd, yd], color="#9a9a9a",
                    linewidth=1.4, zorder=0)

    # column family headers + bracket lines
    for t, start, n in col_group_spans:
        xmid = col_x(start) + n * CELL_W / 2
        ax.text(xmid, -3.30, t, ha="center", va="bottom", fontsize=8.5,
                fontweight="bold", color="#555555")
        ax.plot([col_x(start) + 0.08, col_x(start) + n * CELL_W - 0.08],
                [-3.12, -3.12], color=MUTED, linewidth=0.9)

    # column labels (rotated, aggregates bold)
    for ci, (short, label, agg) in enumerate(flat_cols):
        ax.text(col_x(ci) + CELL_W / 2, -0.12, label, ha="left", va="bottom",
                rotation=52, rotation_mode="anchor", fontsize=8, color=DARK,
                fontweight="bold" if agg else "normal")

    # subtle separator above any within-prompt row (it uses a different metric)
    for row_id, plabel, metric, yc in placed:
        if metric == "zw":
            ax.plot([-0.02, total_w + 0.02], [yc - 0.02, yc - 0.02],
                    color="#b0b0b0", linewidth=0.9, linestyle=(0, (3, 2)), zorder=0)

    # row labels + per-group rotated labels
    for row_id, plabel, metric, yc in placed:
        ax.text(-0.18, yc + CELL_H / 2, plabel, ha="right", va="center",
                fontsize=8.5, fontstyle=("italic" if metric == "zw" else "normal"),
                color=("#1a6fb0" if metric == "zw" else "#333333"))
    for gname, gs, ge in spans:
        ax.text(-3.05, (gs + ge) / 2, gname, ha="center", va="center", rotation=90,
                fontsize=8.5, fontweight="bold", color="#777777")
        ax.plot([-2.7, -2.7], [gs + 0.04, ge - 0.04], color=MUTED, linewidth=1.0)

    # legend
    leg_y = total_h + 1.1
    ax.text(0, leg_y - 0.6, "Legend", ha="left", va="center", fontsize=9.5,
            fontweight="bold", color="#333333")
    for i, r in enumerate([-0.6, -0.3, 0.0, 0.3, 0.6]):
        lx = i * 2.3
        ax.add_patch(FancyBboxPatch((lx, leg_y), 0.62, 0.42,
                     boxstyle="round,pad=0,rounding_size=0.05",
                     facecolor=cell_rgb(r), edgecolor="#dddddd", linewidth=0.7))
        ax.text(lx + 0.78, leg_y + 0.21, f"ρ = {r:+.1f}".replace("+0.0", "0"),
                ha="left", va="center", fontsize=8, color="#444444")
    ax.text(0, leg_y + 1.0, "*** p<0.001     ** p<0.01     * p<0.05",
            ha="left", va="center", fontsize=8, color="#888888")

    ax.set_xlim(-3.5, total_w + 0.15)
    ax.set_ylim(leg_y + 1.4, -4.3)
    ax.axis("off")
    ax.set_aspect("equal")
    ax.set_title(title, fontsize=12, fontweight="bold", pad=18, loc="left", x=0.0)
    fig.savefig(outfile, dpi=170, bbox_inches="tight", facecolor="white")
    plt.close(fig)
    print(f"wrote {outfile}")


render(FIG_ROWS, "correlation_heatmap.png",
       "Correlation with downstream benchmarks (raw-pooled; last row is within-prompt)")