Datasets:

penfever
/

jupiter-tasktrove-dapo-artifacts

#!/usr/bin/env python3
"""Extract per-step training curves for the jupiter-tasktrove-dapo arms (D0-D4) and plot them.
Adapted from ``experiments/complete/tasktrove-hparam-optimization/artifacts/plot_reward_curves.py``.
Differences: Jupiter only (these arms never ran on Iris), no HF gap filler, and the DAPO-specific
metrics (dynamic-sampling discard rate, Section 3.4 overlong incidence/penalty, shaped reward) get
their own panels.
Source
------
Slurm ``.out`` logs under ``/e/data1/datasets/playground/ot-baf/jtd-d<N>*/logs/``, reached over ssh.
Every log carries ``WANDB_MIRROR kind=train step=N metrics={...}`` lines (Python repr, single
quotes). One extractor runs on Jupiter and returns JSON; it ships only files whose mtime/size
moved since the cached copy under ``curves/cache/``.
Lineages
--------
An arm's chains resume from a shared checkpoint bank (``jtd-<arm>/jtd-<arm>/checkpoints``), so steps
from several experiment roots (``jtd-d0_8``, ``jtd-d0_9``, ...) form ONE series. Chains that were
restarted from scratch (retired submissions in TRACKER.md) overlap those step numbers and must not
be merged in. ``LINEAGE`` names the roots of each arm's current series; every other root with train
lines is plotted as a faint dashed "retired" line on the per-arm panel and excluded from the
combined comparisons. Within a lineage, the most recently written log wins a step (a resume
re-does the in-flight step).
Reading the plots
-----------------
* ``reward/avg_pass_at_8`` on D1, D2 and D4 is measured AFTER dynamic-sampling's filter, so it is
  high by construction (uninformative all-pass / all-fail groups are discarded). Compare D0/D3 to
  D1/D2/D4 on ``reward/avg_raw_reward`` and on the never-filtered diagnostics, not on pass@8.
* D4 trains on the Harbor ``threshold`` shaped reward; its raw reward is the continuous verifier
  pass-ratio, not the binary verdict the other arms use.
* Compare at matched STEPS. D0 and D3 pay no resampling cost and out-run D1/D2/D4 in wall-clock.
Usage
-----
    python plot_dapo_curves.py                 # extract (cached) + plot into ./curves
    python plot_dapo_curves.py --from-csv curves/dapo_curves.csv
    python plot_dapo_curves.py --refresh       # re-ship every log
"""

from __future__ import annotations

import argparse
import ast
import csv
import json
import os
import re
import shlex
import subprocess
import sys
from collections import defaultdict

BASE = "/e/data1/datasets/playground/ot-baf"
SSH_HOST = "Jupiter"
ARMS = ["d0", "d1", "d2", "d3", "d4", "d5"]

LABELS = {
    "d0": "D0 control: GRPO, eps 0.2/0.2, no dynamic sampling",
    "d1": "D1 DAPO core: clip-higher 0.4 + dynamic sampling",
    "d2": "D2 full DAPO: D1 + Section 3.4 overlong (l_max 12288 / l_cache 3072)",
    "d3": "D3 Section 3.4 only: D0 + overlong (l_max 12288 / l_cache 3072)",
    "d4": "D4 DAPO core + threshold partial credit (1.0 / 0.3), informative_on=shaped",
    "d5": "D5 GSPO: D0 with policy_loss_type=gspo, eps 3e-4/4e-4, loss_reduction=sequence_mean",
}

# Roots of each arm's CURRENT series (TRACKER.md "Arms" table + "Retired submissions").
# D0: 1436043 (jtd-d0_7, post-#422 relaunch) banked steps 1-6; chain 1442713-18 (jtd-d0_8) resumed
#     at 7; the fresh chain 1464927+ (jtd-d0_9) resumes from 42. jtd-d0_2 (1432018) is the retired
#     pre-migration chain.
# D1: 1447018+ (jtd-d1_9) banked steps 1-3/4, 1457840 (jtd-d1_10) and 1461200 (jtd-d1_11) resumed it.
# D2: FRESH chain 1466109+ launched 11:5x UTC 08-23 into a clean `jtd-d2` root with
#     `dynamic_sampling.informative_on: unshaped`. Every earlier D2 root (jtd-d2 old, jtd-d2_2, jtd-d2_3)
#     was purged from Jupiter (logs archived in artifacts/retired/jtd-d2-logs-20260823.tgz); their rows
#     survive only in the CSV/cache and are treated as retired.
# D3: jtd-d3_3 is FRESH at l_max 12288; jtd-d3 (l_max 32768, retired) and jtd-d3_2 (gate-stopped)
#     are earlier parameterizations and overlap its steps.
# D4: 1459494 (jtd-d4) banked steps 1-2, 1461260+ (jtd-d4_2) resumed.
# D5: FRESH chain 1468466+ into `jtd-d5` (launched 17:09 UTC 08-23, after the cancelled 1468452-57).
#     Its links die early and often (6/6 on the NCCL-abort family, escalated 08-24), so the series is
#     stitched from many short links resuming off the shared checkpoint bank; steps are still one series.
LINEAGE = {
    "d0": ["jtd-d0_7", "jtd-d0_8", "jtd-d0_9"],
    "d1": ["jtd-d1_9", "jtd-d1_10", "jtd-d1_11", "jtd-d1_12"],
    "d2": ["jtd-d2"],
    "d3": ["jtd-d3_3", "jtd-d3_4"],
    "d4": ["jtd-d4", "jtd-d4_2", "jtd-d4_3", "jtd-d4_4"],
    "d5": ["jtd-d5"],
}

METRICS = [
    ("reward", "reward/avg_raw_reward"),
    ("pass_at_8", "reward/avg_pass_at_8"),
    ("entropy", "policy/policy_entropy"),
    ("grad_norm", "policy/raw_grad_norm"),
    ("adv_abs", "loss/avg_raw_advantages_abs"),
    ("tokens", "generate/avg_num_tokens"),
    ("discard_rate", "async/dynamic_sampling/discarded_rate"),
    ("overlong_incidence", "generate/reward_shaping/overlong_incidence"),
    ("overlong_penalty", "generate/reward_shaping/overlong_penalty_mean"),
    ("shaped_reward", "generate/reward_shaping/shaped_reward_mean"),
    ("response_tokens", "generate/reward_shaping/response_tokens_mean"),
    ("clip_high", "policy/ppo_clip_ratio_high"),
    ("clip_low", "policy/ppo_clip_ratio_low"),
]
COLS = [c for c, _ in METRICS]
CSV_COLS = ["arm", "root", "job", "mtime", "step", *COLS]

ROOT_RE = re.compile(r"^jtd-(d\d)(?:_\d+)?$")
TRAIN_LINE_RE = re.compile(r"step=(\d+) metrics=(\{.*\})")
ANSI_RE = re.compile(r"\x1b\[[0-9;]*m")

# Runs on Jupiter. argv[1]=BASE, argv[2]=JSON {path: [mtime, size]} already cached.
# Emits JSON {path: {"mtime":..,"size":..,"lines":[raw WANDB_MIRROR lines]}} for changed files.
EXTRACTOR = r'''
import glob, json, os, sys
base, known = sys.argv[1], json.loads(sys.argv[2])
out = {}
for path in sorted(glob.glob(os.path.join(base, "jtd-d[0-9]*", "logs", "*.out"))):
    if "__dryrun" in path:
        continue
    st = os.stat(path)
    if known.get(path) == [int(st.st_mtime), st.st_size]:
        continue
    lines = []
    with open(path, "rb") as fh:
        for raw in fh:
            if b"WANDB_MIRROR kind=train" in raw:
                lines.append(raw.decode("utf-8", errors="replace").rstrip("\n"))
    if lines:
        out[path] = {"mtime": int(st.st_mtime), "size": st.st_size, "lines": lines}
json.dump(out, sys.stdout)
'''


def parse_line(line: str) -> tuple[int, dict] | None:
    m = TRAIN_LINE_RE.search(ANSI_RE.sub("", line))
    if not m:
        return None
    try:
        d = json.loads(m.group(2))
    except ValueError:
        try:
            d = ast.literal_eval(m.group(2))
        except (ValueError, SyntaxError):
            return None
    return int(m.group(1)), d


def extract(cache_path: str, refresh: bool) -> dict:
    """Return {path: {mtime, size, rows: {step: {col: val}}}} for every Jupiter log with train lines."""
    cache = {}
    if os.path.exists(cache_path) and not refresh:
        with open(cache_path) as f:
            cache = json.load(f)
    known = {p: [v["mtime"], v["size"]] for p, v in cache.items()}
    cmd = f"python3 - {shlex.quote(BASE)} {shlex.quote(json.dumps(known))}"
    proc = subprocess.run(["ssh", SSH_HOST, cmd], input=EXTRACTOR, capture_output=True, text=True,
                          check=True)
    fresh = json.loads(proc.stdout)
    for path, rec in fresh.items():
        rows: dict[str, dict] = {}
        for line in rec["lines"]:
            parsed = parse_line(line)
            if parsed is None:
                continue
            step, d = parsed
            rows[str(step)] = {col: d.get(key) for col, key in METRICS}
        cache[path] = {"mtime": rec["mtime"], "size": rec["size"], "rows": rows}
    os.makedirs(os.path.dirname(cache_path), exist_ok=True)
    with open(cache_path, "w") as f:
        json.dump(cache, f)
    print(f"  {len(fresh)} log(s) re-shipped, {len(cache)} in cache")
    return cache


def rows_from_cache(cache: dict) -> list[dict]:
    out = []
    for path, rec in cache.items():
        root = os.path.basename(os.path.dirname(os.path.dirname(path)))
        m = ROOT_RE.match(root)
        if not m:
            continue
        job = os.path.basename(path).rsplit("_", 1)[-1].removesuffix(".out")
        for step, vals in rec["rows"].items():
            out.append({"arm": m.group(1), "root": root, "job": job, "mtime": rec["mtime"],
                        "step": int(step), **vals})
    return out


def write_csv(path: str, rows: list[dict]) -> None:
    with open(path, "w", newline="") as f:
        w = csv.DictWriter(f, fieldnames=CSV_COLS)
        w.writeheader()
        for r in sorted(rows, key=lambda r: (r["arm"], r["root"], r["step"], r["mtime"])):
            w.writerow(r)


def read_csv(path: str) -> list[dict]:
    with open(path) as f:
        rows = []
        for r in csv.DictReader(f):
            r["mtime"] = int(r["mtime"])
            r["step"] = int(r["step"])
            for c in COLS:
                r[c] = float(r[c]) if r[c] not in ("", "None") else None
            rows.append(r)
    return rows


def series(rows: list[dict]) -> tuple[dict, dict]:
    """Split rows into (current, retired): arm -> {step: row}. Latest mtime wins within a lineage;
    retired roots are keyed as "<arm>/<root>" so each retired chain keeps its own line."""
    current: dict[str, dict[int, dict]] = defaultdict(dict)
    retired: dict[str, dict[int, dict]] = defaultdict(dict)
    for r in sorted(rows, key=lambda r: r["mtime"]):
        if r["root"] in LINEAGE.get(r["arm"], []):
            current[r["arm"]][r["step"]] = r
        else:
            retired[f"{r['arm']}/{r['root']}"][r["step"]] = r
    return current, retired


GAP_BREAK = 5


def segments(pts: list[tuple], max_gap: int = GAP_BREAK) -> list[list[tuple]]:
    runs: list[list[tuple]] = []
    for pt in pts:
        if runs and pt[0] - runs[-1][-1][0] <= max_gap:
            runs[-1].append(pt)
        else:
            runs.append([pt])
    return runs


def ema(vals: list[float], window: int = 5) -> list[float]:
    out = []
    for i in range(len(vals)):
        lo = max(0, i - window + 1)
        out.append(sum(vals[lo:i + 1]) / len(vals[lo:i + 1]))
    return out


def points(s: dict[int, dict], col: str) -> list[tuple[int, float]]:
    return [(step, r[col]) for step, r in sorted(s.items()) if r.get(col) is not None]


ARM_COLORS = {"d0": "black", "d1": "red", "d2": "blue", "d3": "green", "d4": "fuchsia", "d5": "darkorange"}

COMBINED = [
    ("reward", "reward / avg_raw_reward", "raw reward (D4: continuous pass-ratio, others: binary)"),
    ("pass_at_8", "reward / avg_pass_at_8", "pass@8 (POST-FILTER on D1/D2/D4 -- high by construction)"),
    ("entropy", "policy / policy_entropy", "policy entropy"),
    ("grad_norm", "policy / raw_grad_norm", "raw grad norm"),
    ("tokens", "generate / avg_num_tokens", "mean generated tokens per trajectory"),
    ("discard_rate", "dynamic_sampling / discarded_rate", "dynamic-sampling discard rate (D1/D2/D4 only)"),
    ("overlong_incidence", "reward_shaping / overlong_incidence", "Section 3.4 overlong incidence (D2/D3)"),
]


def _line(ax, pts, *, color, label, smooth, ls="-", alpha=1.0, lw=2.0, marker=None):
    first = True
    for run in segments(pts):
        xs = [x for x, _ in run]
        ys = [y for _, y in run]
        ax.plot(xs, ema(ys) if smooth else ys, color=color, ls=ls, lw=lw, alpha=alpha,
                marker=marker, ms=3, label=label if first else None)
        first = False


def plot_combined(plt, current: dict, out_dir: str) -> None:
    for col, ylabel, title in COMBINED:
        fig, ax = plt.subplots(figsize=(12, 5.6))
        any_pts = False
        for arm in ARMS:
            pts = points(current.get(arm, {}), col)
            if len(pts) < 2:
                continue
            any_pts = True
            _line(ax, pts, color=ARM_COLORS[arm], label=f"{arm.upper()} (n={len(pts)})", smooth=False,
                  alpha=0.35, lw=1.0, marker="o")
            _line(ax, pts, color=ARM_COLORS[arm], label=None, smooth=True, lw=2.2)
        if not any_pts:
            plt.close(fig)
            continue
        ax.set_xlabel("training step")
        ax.set_ylabel(ylabel)
        ax.set_title(f"jupiter-tasktrove-dapo -- {title}; faint = raw, bold = trailing-5 EMA")
        ax.grid(True, alpha=0.3)
        ax.legend(fontsize=9)
        fig.tight_layout()
        fig.savefig(os.path.join(out_dir, f"combined_{col}.png"), dpi=130)
        plt.close(fig)


PANEL = [
    ("reward", "raw reward"),
    ("pass_at_8", "pass@8"),
    ("entropy", "entropy"),
    ("grad_norm", "grad norm"),
    ("tokens", "gen tokens"),
    ("discard_rate", "discard rate"),
    ("overlong_incidence", "overlong incidence"),
    ("overlong_penalty", "overlong penalty"),
    ("shaped_reward", "shaped reward"),
]


def plot_arm(plt, arm: str, cur: dict[int, dict], retired: dict[str, dict[int, dict]], out_dir: str) -> None:
    panels = [(col, name) for col, name in PANEL if len(points(cur, col)) >= 1
              or any(points(s, col) for s in retired.values())]
    if not panels:
        return
    ncol = 3
    nrow = (len(panels) + ncol - 1) // ncol
    fig, axes = plt.subplots(nrow, ncol, figsize=(15, 3.6 * nrow), squeeze=False)
    for ax, (col, name) in zip(axes.flat, panels):
        for key, s in sorted(retired.items()):
            pts = points(s, col)
            if pts:
                _line(ax, pts, color="0.6", ls="--", lw=1.2, alpha=0.8, smooth=False,
                      label=f"retired {key.split('/', 1)[1]}")
        pts = points(cur, col)
        if pts:
            _line(ax, pts, color=ARM_COLORS[arm], alpha=0.35, lw=1.0, marker="o", smooth=False, label="raw")
            _line(ax, pts, color=ARM_COLORS[arm], lw=2.2, smooth=True, label="EMA-5")
        ax.set_title(name, fontsize=10)
        ax.grid(True, alpha=0.3)
        ax.legend(fontsize=7)
    for ax in list(axes.flat)[len(panels):]:
        ax.axis("off")
    steps = sorted(cur)
    span = f"steps {steps[0]}..{steps[-1]} (n={len(steps)})" if steps else "no current-series steps"
    fig.suptitle(f"{LABELS[arm]}\n{span}; roots {', '.join(LINEAGE[arm])}", fontsize=11)
    fig.tight_layout()
    fig.savefig(os.path.join(out_dir, f"arm_{arm}.png"), dpi=130)
    plt.close(fig)
    print(f"  {arm}: {span}; retired chains: {len(retired)}")


def plot(rows: list[dict], out_dir: str) -> None:
    import matplotlib

    matplotlib.use("Agg")
    import matplotlib.pyplot as plt

    os.makedirs(out_dir, exist_ok=True)
    current, retired = series(rows)
    plot_combined(plt, current, out_dir)
    for arm in ARMS:
        ret = {k: v for k, v in retired.items() if k.startswith(f"{arm}/")}
        plot_arm(plt, arm, current.get(arm, {}), ret, out_dir)
    print(f"wrote plots to {out_dir}")


def main() -> None:
    here = os.path.dirname(os.path.abspath(__file__))
    p = argparse.ArgumentParser(description=__doc__.split("\n")[0])
    p.add_argument("--out-dir", default=os.path.join(here, "curves"))
    p.add_argument("--from-csv", help="re-plot from a CSV written earlier; no ssh")
    p.add_argument("--refresh", action="store_true", help="ignore the mtime/size cache and re-ship every log")
    args = p.parse_args()

    os.makedirs(args.out_dir, exist_ok=True)
    csv_path = os.path.join(args.out_dir, "dapo_curves.csv")
    if args.from_csv:
        rows = read_csv(args.from_csv)
    else:
        print(f"extracting from {SSH_HOST}:{BASE} ...")
        cache = extract(os.path.join(args.out_dir, "cache", "jupiter.json"), args.refresh)
        rows = rows_from_cache(cache)
        write_csv(csv_path, rows)
        print(f"  wrote {csv_path} ({len(rows)} rows)")
    plot(rows, args.out_dir)


if __name__ == "__main__":
    main()

	#!/usr/bin/env python3
	"""Extract per-step training curves for the jupiter-tasktrove-dapo arms (D0-D4) and plot them.

	Adapted from ``experiments/complete/tasktrove-hparam-optimization/artifacts/plot_reward_curves.py``.
	Differences: Jupiter only (these arms never ran on Iris), no HF gap filler, and the DAPO-specific
	metrics (dynamic-sampling discard rate, Section 3.4 overlong incidence/penalty, shaped reward) get
	their own panels.

	Source
	------
	Slurm ``.out`` logs under ``/e/data1/datasets/playground/ot-baf/jtd-d<N>*/logs/``, reached over ssh.
	Every log carries ``WANDB_MIRROR kind=train step=N metrics={...}`` lines (Python repr, single
	quotes). One extractor runs on Jupiter and returns JSON; it ships only files whose mtime/size
	moved since the cached copy under ``curves/cache/``.

	Lineages
	--------
	An arm's chains resume from a shared checkpoint bank (``jtd-<arm>/jtd-<arm>/checkpoints``), so steps
	from several experiment roots (``jtd-d0_8``, ``jtd-d0_9``, ...) form ONE series. Chains that were
	restarted from scratch (retired submissions in TRACKER.md) overlap those step numbers and must not
	be merged in. ``LINEAGE`` names the roots of each arm's current series; every other root with train
	lines is plotted as a faint dashed "retired" line on the per-arm panel and excluded from the
	combined comparisons. Within a lineage, the most recently written log wins a step (a resume
	re-does the in-flight step).

	Reading the plots
	-----------------
	* ``reward/avg_pass_at_8`` on D1, D2 and D4 is measured AFTER dynamic-sampling's filter, so it is
	high by construction (uninformative all-pass / all-fail groups are discarded). Compare D0/D3 to
	D1/D2/D4 on ``reward/avg_raw_reward`` and on the never-filtered diagnostics, not on pass@8.
	* D4 trains on the Harbor ``threshold`` shaped reward; its raw reward is the continuous verifier
	pass-ratio, not the binary verdict the other arms use.
	* Compare at matched STEPS. D0 and D3 pay no resampling cost and out-run D1/D2/D4 in wall-clock.

	Usage
	-----
	python plot_dapo_curves.py # extract (cached) + plot into ./curves
	python plot_dapo_curves.py --from-csv curves/dapo_curves.csv
	python plot_dapo_curves.py --refresh # re-ship every log
	"""

	from __future__ import annotations

	import argparse
	import ast
	import csv
	import json
	import os
	import re
	import shlex
	import subprocess
	import sys
	from collections import defaultdict

	BASE = "/e/data1/datasets/playground/ot-baf"
	SSH_HOST = "Jupiter"
	ARMS = ["d0", "d1", "d2", "d3", "d4", "d5"]

	LABELS = {
	"d0": "D0 control: GRPO, eps 0.2/0.2, no dynamic sampling",
	"d1": "D1 DAPO core: clip-higher 0.4 + dynamic sampling",
	"d2": "D2 full DAPO: D1 + Section 3.4 overlong (l_max 12288 / l_cache 3072)",
	"d3": "D3 Section 3.4 only: D0 + overlong (l_max 12288 / l_cache 3072)",
	"d4": "D4 DAPO core + threshold partial credit (1.0 / 0.3), informative_on=shaped",
	"d5": "D5 GSPO: D0 with policy_loss_type=gspo, eps 3e-4/4e-4, loss_reduction=sequence_mean",
	}

	# Roots of each arm's CURRENT series (TRACKER.md "Arms" table + "Retired submissions").
	# D0: 1436043 (jtd-d0_7, post-#422 relaunch) banked steps 1-6; chain 1442713-18 (jtd-d0_8) resumed
	# at 7; the fresh chain 1464927+ (jtd-d0_9) resumes from 42. jtd-d0_2 (1432018) is the retired
	# pre-migration chain.
	# D1: 1447018+ (jtd-d1_9) banked steps 1-3/4, 1457840 (jtd-d1_10) and 1461200 (jtd-d1_11) resumed it.
	# D2: FRESH chain 1466109+ launched 11:5x UTC 08-23 into a clean `jtd-d2` root with
	# `dynamic_sampling.informative_on: unshaped`. Every earlier D2 root (jtd-d2 old, jtd-d2_2, jtd-d2_3)
	# was purged from Jupiter (logs archived in artifacts/retired/jtd-d2-logs-20260823.tgz); their rows
	# survive only in the CSV/cache and are treated as retired.
	# D3: jtd-d3_3 is FRESH at l_max 12288; jtd-d3 (l_max 32768, retired) and jtd-d3_2 (gate-stopped)
	# are earlier parameterizations and overlap its steps.
	# D4: 1459494 (jtd-d4) banked steps 1-2, 1461260+ (jtd-d4_2) resumed.
	# D5: FRESH chain 1468466+ into `jtd-d5` (launched 17:09 UTC 08-23, after the cancelled 1468452-57).
	# Its links die early and often (6/6 on the NCCL-abort family, escalated 08-24), so the series is
	# stitched from many short links resuming off the shared checkpoint bank; steps are still one series.
	LINEAGE = {
	"d0": ["jtd-d0_7", "jtd-d0_8", "jtd-d0_9"],
	"d1": ["jtd-d1_9", "jtd-d1_10", "jtd-d1_11", "jtd-d1_12"],
	"d2": ["jtd-d2"],
	"d3": ["jtd-d3_3", "jtd-d3_4"],
	"d4": ["jtd-d4", "jtd-d4_2", "jtd-d4_3", "jtd-d4_4"],
	"d5": ["jtd-d5"],
	}

	METRICS = [
	("reward", "reward/avg_raw_reward"),
	("pass_at_8", "reward/avg_pass_at_8"),
	("entropy", "policy/policy_entropy"),
	("grad_norm", "policy/raw_grad_norm"),
	("adv_abs", "loss/avg_raw_advantages_abs"),
	("tokens", "generate/avg_num_tokens"),
	("discard_rate", "async/dynamic_sampling/discarded_rate"),
	("overlong_incidence", "generate/reward_shaping/overlong_incidence"),
	("overlong_penalty", "generate/reward_shaping/overlong_penalty_mean"),
	("shaped_reward", "generate/reward_shaping/shaped_reward_mean"),
	("response_tokens", "generate/reward_shaping/response_tokens_mean"),
	("clip_high", "policy/ppo_clip_ratio_high"),
	("clip_low", "policy/ppo_clip_ratio_low"),
	]
	COLS = [c for c, _ in METRICS]
	CSV_COLS = ["arm", "root", "job", "mtime", "step", *COLS]

	ROOT_RE = re.compile(r"^jtd-(d\d)(?:_\d+)?$")
	TRAIN_LINE_RE = re.compile(r"step=(\d+) metrics=(\{.*\})")
	ANSI_RE = re.compile(r"\x1b\[[0-9;]*m")

	# Runs on Jupiter. argv[1]=BASE, argv[2]=JSON {path: [mtime, size]} already cached.
	# Emits JSON {path: {"mtime":..,"size":..,"lines":[raw WANDB_MIRROR lines]}} for changed files.
	EXTRACTOR = r'''
	import glob, json, os, sys
	base, known = sys.argv[1], json.loads(sys.argv[2])
	out = {}
	for path in sorted(glob.glob(os.path.join(base, "jtd-d[0-9]", "logs", ".out"))):
	if "__dryrun" in path:
	continue
	st = os.stat(path)
	if known.get(path) == [int(st.st_mtime), st.st_size]:
	continue
	lines = []
	with open(path, "rb") as fh:
	for raw in fh:
	if b"WANDB_MIRROR kind=train" in raw:
	lines.append(raw.decode("utf-8", errors="replace").rstrip("\n"))
	if lines:
	out[path] = {"mtime": int(st.st_mtime), "size": st.st_size, "lines": lines}
	json.dump(out, sys.stdout)
	'''


	def parse_line(line: str) -> tuple[int, dict] \| None:
	m = TRAIN_LINE_RE.search(ANSI_RE.sub("", line))
	if not m:
	return None
	try:
	d = json.loads(m.group(2))
	except ValueError:
	try:
	d = ast.literal_eval(m.group(2))
	except (ValueError, SyntaxError):
	return None
	return int(m.group(1)), d


	def extract(cache_path: str, refresh: bool) -> dict:
	"""Return {path: {mtime, size, rows: {step: {col: val}}}} for every Jupiter log with train lines."""
	cache = {}
	if os.path.exists(cache_path) and not refresh:
	with open(cache_path) as f:
	cache = json.load(f)
	known = {p: [v["mtime"], v["size"]] for p, v in cache.items()}
	cmd = f"python3 - {shlex.quote(BASE)} {shlex.quote(json.dumps(known))}"
	proc = subprocess.run(["ssh", SSH_HOST, cmd], input=EXTRACTOR, capture_output=True, text=True,
	check=True)
	fresh = json.loads(proc.stdout)
	for path, rec in fresh.items():
	rows: dict[str, dict] = {}
	for line in rec["lines"]:
	parsed = parse_line(line)
	if parsed is None:
	continue
	step, d = parsed
	rows[str(step)] = {col: d.get(key) for col, key in METRICS}
	cache[path] = {"mtime": rec["mtime"], "size": rec["size"], "rows": rows}
	os.makedirs(os.path.dirname(cache_path), exist_ok=True)
	with open(cache_path, "w") as f:
	json.dump(cache, f)
	print(f" {len(fresh)} log(s) re-shipped, {len(cache)} in cache")
	return cache


	def rows_from_cache(cache: dict) -> list[dict]:
	out = []
	for path, rec in cache.items():
	root = os.path.basename(os.path.dirname(os.path.dirname(path)))
	m = ROOT_RE.match(root)
	if not m:
	continue
	job = os.path.basename(path).rsplit("_", 1)[-1].removesuffix(".out")
	for step, vals in rec["rows"].items():
	out.append({"arm": m.group(1), "root": root, "job": job, "mtime": rec["mtime"],
	"step": int(step), **vals})
	return out


	def write_csv(path: str, rows: list[dict]) -> None:
	with open(path, "w", newline="") as f:
	w = csv.DictWriter(f, fieldnames=CSV_COLS)
	w.writeheader()
	for r in sorted(rows, key=lambda r: (r["arm"], r["root"], r["step"], r["mtime"])):
	w.writerow(r)


	def read_csv(path: str) -> list[dict]:
	with open(path) as f:
	rows = []
	for r in csv.DictReader(f):
	r["mtime"] = int(r["mtime"])
	r["step"] = int(r["step"])
	for c in COLS:
	r[c] = float(r[c]) if r[c] not in ("", "None") else None
	rows.append(r)
	return rows


	def series(rows: list[dict]) -> tuple[dict, dict]:
	"""Split rows into (current, retired): arm -> {step: row}. Latest mtime wins within a lineage;
	retired roots are keyed as "<arm>/<root>" so each retired chain keeps its own line."""
	current: dict[str, dict[int, dict]] = defaultdict(dict)
	retired: dict[str, dict[int, dict]] = defaultdict(dict)
	for r in sorted(rows, key=lambda r: r["mtime"]):
	if r["root"] in LINEAGE.get(r["arm"], []):
	current[r["arm"]][r["step"]] = r
	else:
	retired[f"{r['arm']}/{r['root']}"][r["step"]] = r
	return current, retired


	GAP_BREAK = 5


	def segments(pts: list[tuple], max_gap: int = GAP_BREAK) -> list[list[tuple]]:
	runs: list[list[tuple]] = []
	for pt in pts:
	if runs and pt[0] - runs[-1][-1][0] <= max_gap:
	runs[-1].append(pt)
	else:
	runs.append([pt])
	return runs


	def ema(vals: list[float], window: int = 5) -> list[float]:
	out = []
	for i in range(len(vals)):
	lo = max(0, i - window + 1)
	out.append(sum(vals[lo:i + 1]) / len(vals[lo:i + 1]))
	return out


	def points(s: dict[int, dict], col: str) -> list[tuple[int, float]]:
	return [(step, r[col]) for step, r in sorted(s.items()) if r.get(col) is not None]


	ARM_COLORS = {"d0": "black", "d1": "red", "d2": "blue", "d3": "green", "d4": "fuchsia", "d5": "darkorange"}

	COMBINED = [
	("reward", "reward / avg_raw_reward", "raw reward (D4: continuous pass-ratio, others: binary)"),
	("pass_at_8", "reward / avg_pass_at_8", "pass@8 (POST-FILTER on D1/D2/D4 -- high by construction)"),
	("entropy", "policy / policy_entropy", "policy entropy"),
	("grad_norm", "policy / raw_grad_norm", "raw grad norm"),
	("tokens", "generate / avg_num_tokens", "mean generated tokens per trajectory"),
	("discard_rate", "dynamic_sampling / discarded_rate", "dynamic-sampling discard rate (D1/D2/D4 only)"),
	("overlong_incidence", "reward_shaping / overlong_incidence", "Section 3.4 overlong incidence (D2/D3)"),
	]


	def _line(ax, pts, *, color, label, smooth, ls="-", alpha=1.0, lw=2.0, marker=None):
	first = True
	for run in segments(pts):
	xs = [x for x, _ in run]
	ys = [y for _, y in run]
	ax.plot(xs, ema(ys) if smooth else ys, color=color, ls=ls, lw=lw, alpha=alpha,
	marker=marker, ms=3, label=label if first else None)
	first = False


	def plot_combined(plt, current: dict, out_dir: str) -> None:
	for col, ylabel, title in COMBINED:
	fig, ax = plt.subplots(figsize=(12, 5.6))
	any_pts = False
	for arm in ARMS:
	pts = points(current.get(arm, {}), col)
	if len(pts) < 2:
	continue
	any_pts = True
	_line(ax, pts, color=ARM_COLORS[arm], label=f"{arm.upper()} (n={len(pts)})", smooth=False,
	alpha=0.35, lw=1.0, marker="o")
	_line(ax, pts, color=ARM_COLORS[arm], label=None, smooth=True, lw=2.2)
	if not any_pts:
	plt.close(fig)
	continue
	ax.set_xlabel("training step")
	ax.set_ylabel(ylabel)
	ax.set_title(f"jupiter-tasktrove-dapo -- {title}; faint = raw, bold = trailing-5 EMA")
	ax.grid(True, alpha=0.3)
	ax.legend(fontsize=9)
	fig.tight_layout()
	fig.savefig(os.path.join(out_dir, f"combined_{col}.png"), dpi=130)
	plt.close(fig)


	PANEL = [
	("reward", "raw reward"),
	("pass_at_8", "pass@8"),
	("entropy", "entropy"),
	("grad_norm", "grad norm"),
	("tokens", "gen tokens"),
	("discard_rate", "discard rate"),
	("overlong_incidence", "overlong incidence"),
	("overlong_penalty", "overlong penalty"),
	("shaped_reward", "shaped reward"),
	]


	def plot_arm(plt, arm: str, cur: dict[int, dict], retired: dict[str, dict[int, dict]], out_dir: str) -> None:
	panels = [(col, name) for col, name in PANEL if len(points(cur, col)) >= 1
	or any(points(s, col) for s in retired.values())]
	if not panels:
	return
	ncol = 3
	nrow = (len(panels) + ncol - 1) // ncol
	fig, axes = plt.subplots(nrow, ncol, figsize=(15, 3.6 * nrow), squeeze=False)
	for ax, (col, name) in zip(axes.flat, panels):
	for key, s in sorted(retired.items()):
	pts = points(s, col)
	if pts:
	_line(ax, pts, color="0.6", ls="--", lw=1.2, alpha=0.8, smooth=False,
	label=f"retired {key.split('/', 1)[1]}")
	pts = points(cur, col)
	if pts:
	_line(ax, pts, color=ARM_COLORS[arm], alpha=0.35, lw=1.0, marker="o", smooth=False, label="raw")
	_line(ax, pts, color=ARM_COLORS[arm], lw=2.2, smooth=True, label="EMA-5")
	ax.set_title(name, fontsize=10)
	ax.grid(True, alpha=0.3)
	ax.legend(fontsize=7)
	for ax in list(axes.flat)[len(panels):]:
	ax.axis("off")
	steps = sorted(cur)
	span = f"steps {steps[0]}..{steps[-1]} (n={len(steps)})" if steps else "no current-series steps"
	fig.suptitle(f"{LABELS[arm]}\n{span}; roots {', '.join(LINEAGE[arm])}", fontsize=11)
	fig.tight_layout()
	fig.savefig(os.path.join(out_dir, f"arm_{arm}.png"), dpi=130)
	plt.close(fig)
	print(f" {arm}: {span}; retired chains: {len(retired)}")


	def plot(rows: list[dict], out_dir: str) -> None:
	import matplotlib

	matplotlib.use("Agg")
	import matplotlib.pyplot as plt

	os.makedirs(out_dir, exist_ok=True)
	current, retired = series(rows)
	plot_combined(plt, current, out_dir)
	for arm in ARMS:
	ret = {k: v for k, v in retired.items() if k.startswith(f"{arm}/")}
	plot_arm(plt, arm, current.get(arm, {}), ret, out_dir)
	print(f"wrote plots to {out_dir}")


	def main() -> None:
	here = os.path.dirname(os.path.abspath(__file__))
	p = argparse.ArgumentParser(description=__doc__.split("\n")[0])
	p.add_argument("--out-dir", default=os.path.join(here, "curves"))
	p.add_argument("--from-csv", help="re-plot from a CSV written earlier; no ssh")
	p.add_argument("--refresh", action="store_true", help="ignore the mtime/size cache and re-ship every log")
	args = p.parse_args()

	os.makedirs(args.out_dir, exist_ok=True)
	csv_path = os.path.join(args.out_dir, "dapo_curves.csv")
	if args.from_csv:
	rows = read_csv(args.from_csv)
	else:
	print(f"extracting from {SSH_HOST}:{BASE} ...")
	cache = extract(os.path.join(args.out_dir, "cache", "jupiter.json"), args.refresh)
	rows = rows_from_cache(cache)
	write_csv(csv_path, rows)
	print(f" wrote {csv_path} ({len(rows)} rows)")
	plot(rows, args.out_dir)


	if __name__ == "__main__":
	main()