File size: 40,133 Bytes
e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a 5c30113 e0eb79a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 | """Upload the trained MiniHack ReMDM checkpoints and results to the Hugging Face Hub.
Discovers every checkpoint under ``checkpoints/``, every ablation run under
``experiments/rl_finetuning/outputs/``, every ``--mode inference`` result
under ``results/inference/`` and the manuscript figure PDFs under
``results/paper_figures/``, stages them with the repo-relative layout
preserved, drops W&B and hub metadata (which carries the author's account,
hostname and local paths), exports safetensors inference weights alongside each
full training state, generates a model card from the checkpoints' own config
snapshots, and uploads.
HF_TOKEN=hf_xxx uv run python scripts/hf_upload.py \\
--repo-id <ANON_HF_REPO_ID> \\
[--inference-results PATH ...] [--dry-run] [--private]
"""
from __future__ import annotations
import argparse
import json
import os
import re
import shutil
import sys
import tempfile
from pathlib import Path
import yaml
# scripts/ is not a package, so the helper is imported by bare name.
# Python already puts this file's directory on sys.path when the script is
# run directly; the explicit insert is for the file-location loaders the
# tests use (tests/test_config.py), which do not.
_SCRIPTS = Path(__file__).resolve().parent
if str(_SCRIPTS) not in sys.path:
sys.path.insert(0, str(_SCRIPTS))
from _git_provenance import copy_tracked_file # noqa: E402
ROOT = Path(__file__).resolve().parent.parent
CKPTS = ROOT / "checkpoints"
RUNS = ROOT / "experiments" / "rl_finetuning" / "outputs"
INFERENCE = ROOT / "results" / "inference"
PAPER_FIGURES = ROOT / "results" / "paper_figures"
PAPER = (
"Return-Weighted ELBO Fine-Tuning Degrades "
"Masked Diffusion Planners"
)
CODE_URL = "https://github.com/ANONYMOUS/remdm-planners"
ENV_NAME = "MiniHack"
ROLES = {
"offline": "Diffusion planner (offline BC)",
"online": "Diffusion planner (online DAgger)",
"imported": "Imported baseline",
}
# An ablation run is published as its own summary plus tables and figures; the
# raw per-iteration logs stay in the code repository.
RUN_FILES = ("results.json", "diagnosis.md")
# gdelta/ holds the per-seed gradient measurements behind the paper's
# decomposition appendix; they land inside the run directory, so
# discover_runs() finds them with no further change.
RUN_DIRS = ("tables", "figures", "gdelta")
# Environment provenance, never needed to restore a checkpoint, and dropped
# from every published config in both repos: the nested `_wandb` blob (email,
# host, git remote, absolute paths), every `wandb_*` and `hub_*` key, and
# `use_wandb` itself — which matches neither prefix, so both repos shipped it
# in the released config while claiming to scrub W&B settings. Keys are
# compared lower-cased, because craftax records them UPPERCASE and minihack
# lower-case.
DROP_PREFIXES = ("hub_",)
# `wandb_` is matched anywhere in the key, not only at the start. A prefix
# rule caught `wandb_project` but not `baselines_wandb_project`, which was
# published in both released config.yaml files, nor the sibling's
# `RESUME_WANDB_RUN_ID` -- the same way `use_wandb` slipped past originally.
# The underscore is what keeps it narrow: `wandbish` contains `wandb` but not
# `wandb_`, so the negative assertions still hold in both repos.
DROP_SUBSTRINGS = ("wandb_",)
DROP_KEYS = ("_wandb", "use_wandb")
def is_environment_key(key: str) -> bool:
"""True for a config key that is provenance rather than recipe."""
lowered = key.lower()
return (
lowered in DROP_KEYS
or lowered.startswith(DROP_PREFIXES)
or any(mark in lowered for mark in DROP_SUBSTRINGS)
)
COPY_IGNORE = shutil.ignore_patterns(
".DS_Store", "__pycache__", "*.pyc", "wandb-metadata.json",
)
HUB_IGNORE = ["**/.DS_Store", "**/__pycache__/**", "**/wandb-metadata.json"]
# =============================================================================
# Discovery
# =============================================================================
# `checkpoints/hf/` is where a Hub *download* lands. Publishing from it would
# re-upload already-published artefacts into a nested `checkpoints/hf/...` tree
# on the Hub, so it is never a publish source in either repo.
HF_DOWNLOAD_DIR = "hf"
def _is_download_copy(path: Path) -> bool:
"""True for anything under ``checkpoints/hf/``, wherever it sits."""
return HF_DOWNLOAD_DIR in path.relative_to(CKPTS).parts
def discover_checkpoints() -> dict[Path, list[Path]]:
"""Map each checkpoint directory to its ``.pth`` files, oldest first.
Discovery is at the **released layout**, ``checkpoints/<role>/<name>/`` —
the layout the Hub repo mirrors, matching the sibling repo. A recursive
``rglob("*.pth")`` found everything instead, which is wrong in both
directions: measured on this repo's live tree it discovered 10 directories,
**none of them at the released layout** — seven raw ``dagger_<timestamp>/``
run directories, the repository root itself with two loose files, and **two
already-published artefacts under** ``checkpoints/hf/``, which a publish
would have pushed back up into a nested ``checkpoints/hf/checkpoints/...``
tree.
A training run writes to its own directory, so its checkpoints have to be
copied into ``checkpoints/{offline,online}/<name>/`` before publishing. The
README documents that; it is the same requirement the sibling repo has
always had, and it is now stated on both sides rather than one.
Anything under ``checkpoints/hf/`` is skipped as a download copy, explicitly
rather than by depth arithmetic.
Returns:
``{checkpoint directory: [.pth paths, oldest first]}``.
"""
models: dict[Path, list[Path]] = {}
for pth in sorted(CKPTS.glob("*/*/*.pth")):
if _is_download_copy(pth):
continue
models.setdefault(pth.parent, []).append(pth)
return models
def discover_runs() -> list[Path]:
"""Every ablation output directory holding a ``results.json``."""
if not RUNS.is_dir():
return []
return sorted(d for d in RUNS.iterdir() if (d / "results.json").is_file())
def discover_paper_figures() -> list[Path]:
"""The manuscript figures, as vector PDFs.
Built by ``scripts/paper_figures.py``, which lives in the sibling repo
because each figure reads *both* repositories' ablation ``results.json``
and puts Craftax Classic and MiniHack side by side in one figure. The
resulting PDFs therefore describe this repo's results as much as the
sibling's, and both Hub repos publish the same set from
``results/paper_figures/``.
They were previously published by neither: discovery covered only
``checkpoints/``, the ablation run directories and ``results/inference/``,
so ``results/paper_figures/`` matched nothing and the upload said nothing
about it.
"""
if not PAPER_FIGURES.is_dir():
return []
return sorted(PAPER_FIGURES.glob("*.pdf"))
def discover_inference(extra: list[str]) -> list[Path]:
"""Inference result JSONs: the default directory plus any given paths."""
found: list[Path] = []
for source in [INFERENCE, *(Path(p) for p in extra)]:
if source.is_dir():
found.extend(sorted(source.glob("*.json")))
elif source.is_file():
found.append(source)
elif source != INFERENCE:
print(f"No inference results at {source}.", file=sys.stderr)
return list(dict.fromkeys(found))
# =============================================================================
# Helpers
# =============================================================================
def dir_size_mb(path: Path) -> float:
if path.is_file():
return path.stat().st_size / 1_048_576
return sum(f.stat().st_size for f in path.rglob("*") if f.is_file()) / 1_048_576
def human_size(path: Path) -> str:
mb = dir_size_mb(path)
return f"{mb:.0f} MB" if mb >= 1 else f"{max(mb * 1024, 1):.0f} KB"
def plural(n: int, word: str) -> str:
return f"{n} {word}" if n == 1 else f"{n} {word}s"
def shorten_paths(value):
"""Shorten absolute cluster paths anywhere in a staged JSON document."""
if isinstance(value, dict):
return {k: shorten_paths(v) for k, v in value.items()}
if isinstance(value, list):
return [shorten_paths(v) for v in value]
if isinstance(value, str) and value.startswith("/"):
return "/".join(Path(value).parts[-2:])
return value
def environment_key_paths(value, path: str = "") -> list[str]:
"""Every dotted path at which an environment key appears, at any depth.
The reporting counterpart to :func:`drop_environment_keys`, so the
uploader can name what it removed rather than claiming a scrub happened.
Recursion stops at a key that is itself dropped: its whole subtree goes.
"""
found: list[str] = []
if isinstance(value, dict):
for key, sub in value.items():
name = str(key)
here = f"{path}.{name}" if path else name
if is_environment_key(name):
found.append(here)
else:
found.extend(environment_key_paths(sub, here))
elif isinstance(value, (list, tuple)):
for i, sub in enumerate(value):
found.extend(environment_key_paths(sub, f"{path}[{i}]"))
return found
def drop_environment_keys(value):
"""Strip environment keys at **every** depth, not just the top level.
Filtering only the top level was a real defect in the sibling repo: its
path-shortener recursed while its key-filter did not, so W&B project and
entity survived one level down inside a nested config snapshot and were
published, while the uploader reported a successful scrub. This repo's
checkpoints happen to carry no nested config, so the same one-level filter
leaked nothing here -- clean for a reason that was not the code being
right, and one schema change away from leaking silently.
Mapping types are preserved, so an ``OrderedDict`` state dict stays an
``OrderedDict`` and the published checkpoint keeps its original structure.
"""
if isinstance(value, dict):
kept = {
key: drop_environment_keys(sub)
for key, sub in value.items()
if not is_environment_key(str(key))
}
if type(value) is dict:
return kept
try:
return type(value)(kept)
except TypeError:
# A mapping needing constructor arguments (e.g. a defaultdict
# factory). Structure matters less than not shipping the key.
return kept
if isinstance(value, list):
return [drop_environment_keys(sub) for sub in value]
return value
def scrub(cfg: dict) -> dict:
"""Drop the environment keys and shorten absolute cluster paths.
The prefixes alone missed `use_wandb`, which starts with neither, so it
shipped in every released config; the sibling repo dropped a different set
again, and left `USE_WANDB` with it. Both now drop the same one.
Both passes recurse. They used to disagree -- ``shorten_paths`` descended
and the key filter did not -- which is the asymmetry that leaked W&B
settings from a nested config snapshot in the sibling repo.
"""
return drop_environment_keys(shorten_paths(cfg))
#: An absolute POSIX path with at least three components, as it appears
#: inside a JSON string, a Markdown table or a LaTeX table cell.
# The lookbehind anchors the match to an absolute path. Without it the regex
# also matched from the middle of a relative one: `experiments/rl_finetuning/
# analysis/tables.py` in a generated LaTeX comment became
# `experimentsanalysis/tables.py`, because matching could start at the slash
# after `experiments`.
#
# `@` and `+` are in the class because the lookbehind makes a narrow class
# leak rather than merely under-match: with them omitted,
# `/home/user@dom/proj/run/file.txt` fails at the start (`user@dom` is not one
# component) and the lookbehind then refuses to restart at `/proj`, so there
# is no fallback match and the whole absolute path ships unshortened. `@` is
# exactly where an email-like component appears in a home-directory path,
# which is the leak class this scrub exists for.
_ABS_PATH = re.compile(r"(?<![\w.@+\-])/(?:[\w.@+\-]+/){2,}[\w.@+\-]+")
#: What a staged tree holds that is worth reading as text.
_TEXT_SUFFIXES = (".md", ".txt", ".csv", ".tex")
_DOC_SUFFIXES = (".json", ".yaml", ".yml")
def shorten_text_paths(text: str) -> str:
"""Apply the ``shorten_paths`` rule to paths embedded in plain text.
``shorten_paths`` only reaches strings it can find by walking a parsed
document. The same cluster paths appear inside ``diagnosis.md`` prose and
LaTeX table cells, where nothing parses them.
"""
return _ABS_PATH.sub(lambda m: "/".join(Path(m.group(0)).parts[-2:]), text)
def scrub_staged_tree(staging: Path) -> list[str]:
"""Scrub every document and text file in the staged tree, in place.
The stagers copy whole directories -- ``tables/``, ``figures/``,
``gdelta/`` -- so naming the files to scrub means maintaining a list by
hand, and a list of known filenames is the same mistake as a known
nesting. Walking what was actually staged makes the set of files scrubbed
identical to the set published, by construction.
This closes a reach bug, not a depth one: ``stage_runs`` copied
``results.json`` with ``shutil.copy2`` and scrubbed nothing at all, so the
released file carried `wandb_project`, `wandb_entity` and the absolute
path of the machine it ran on. ``scrub`` was correct and was simply never
called on it.
A file is rewritten only when scrubbing changes it, so anything already
clean keeps its bytes.
Returns:
Staged-relative paths of the files that changed.
"""
changed: list[str] = []
for path in sorted(staging.rglob("*")):
if not path.is_file():
continue
suffix = path.suffix.lower()
if suffix in _DOC_SUFFIXES:
raw = path.read_text()
try:
doc = (json.loads(raw) if suffix == ".json"
else yaml.safe_load(raw))
except (json.JSONDecodeError, yaml.YAMLError):
continue
if not isinstance(doc, (dict, list)):
continue
clean = drop_environment_keys(shorten_paths(doc))
if clean == doc:
continue
path.write_text(
json.dumps(clean, indent=2) + "\n" if suffix == ".json"
else yaml.safe_dump(clean, sort_keys=False)
)
changed.append(str(path.relative_to(staging)))
elif suffix in _TEXT_SUFFIXES:
raw = path.read_text()
clean = shorten_text_paths(raw)
if clean != raw:
path.write_text(clean)
changed.append(str(path.relative_to(staging)))
return changed
def scrub_checkpoint(src: Path, dst: Path) -> None:
"""Copy a ``.pth`` with its provenance keys dropped.
``scrub`` only reaches the YAML/JSON sidecars, so `wandb_run_id` shipped
inside the pickled checkpoints themselves (iter563.pth -> runid002,
offline_step50000.pth -> runid003) while the uploader reported that W&B
settings had been scrubbed. Environment keys are dropped with the same
predicate the configs use; every training-state key is preserved.
"""
import torch
ckpt = torch.load(src, map_location="cpu", weights_only=False)
if not isinstance(ckpt, dict):
shutil.copy2(src, dst)
return
dropped = environment_key_paths(ckpt)
if not dropped:
# Nothing to remove, so copy the bytes rather than re-saving them:
# torch.save is not byte-reproducible, and a clean checkpoint should
# not change on the Hub just for having been inspected.
shutil.copy2(src, dst)
return
torch.save(drop_environment_keys(ckpt), dst)
print(f" scrubbed {', '.join(sorted(dropped))} from {src.name}")
def export_weights(pth: Path, out: Path) -> dict:
"""Write EMA inference weights as safetensors; return checkpoint stats."""
import torch
from safetensors.torch import save_file
ckpt = torch.load(pth, map_location="cpu", weights_only=False)
model_sd = ckpt["model_state_dict"]
ema_sd = ckpt.get("ema_state_dict") or {}
source = "ema_state_dict" if set(ema_sd) >= set(model_sd) else "model_state_dict"
tensors = {
k: v.contiguous()
for k, v in (ema_sd if source == "ema_state_dict" else model_sd).items()
}
save_file(tensors, str(out), metadata={"format": "pt", "source": source})
# The offline trainer stores step x batch under `env_steps`; those are
# sample-equivalents, not env.step() calls. Only DAgger's are real.
if "iteration" in ckpt:
counter, value = "dagger_iteration", ckpt["iteration"]
step, unit = f"iteration {value:,}", "env steps"
elif "step" in ckpt:
counter, value = "gradient_step", ckpt["step"]
step, unit = f"gradient step {value:,}", "sample-equivalents"
else:
counter, value = None, None
step, unit = "final", "sample-equivalents"
return {
"params": sum(v.numel() for v in tensors.values()),
"source": source,
"counter": counter,
"value": value,
"step": step,
"detail": f"{ckpt['env_steps']:,} {unit}" if ckpt.get("env_steps") else step,
"keys": sorted(
k for k in ckpt
if k.endswith("state_dict") or k in {"rng_states", "curriculum_state"}
),
}
def required(cfg: dict, key: str) -> object:
"""Read a key the published record depends on, or fail the upload.
``.get()`` here published a null instead: every DAgger `selection.json`
released before this guard recorded `"every": null` because the two keys
read had been renamed out of the config, while the model card claims each
file records the budget it came from. A declared-but-null value is still
accepted -- the four `offline_*` keys are nullable by design.
"""
if key not in cfg:
raise KeyError(
f"{key} is absent from the checkpoint's config snapshot, so "
f"selection.json cannot record the candidate set it came from"
)
return cfg[key]
def selection(cfg: dict, stats: dict, metric: str | None) -> dict:
"""Record the best-of-N selection a published checkpoint came from."""
if stats["counter"] == "gradient_step":
candidates = {
"unit": "gradient_steps",
"every": required(cfg, "offline_checkpoint_every_grad_steps"),
"configured_budget": required(cfg, "offline_total_grad_steps"),
}
else:
# DAgger checkpoints on an env-step cadence (`online.py` compares
# against `env_steps_total`) but names its files by iteration, so the
# candidate set is denominated in env steps and `selected` is not.
candidates = {
"unit": "env_steps",
"every": required(cfg, "checkpoint_every_timesteps"),
"configured_max": required(cfg, "total_timesteps"),
}
return {
"policy": "best-of-N over periodic checkpoints",
"selected": {stats["counter"]: stats["value"]},
"selection_metric": metric,
"candidates": candidates,
"eval_protocol": {
"episodes_per_env": required(cfg, "checkpoint_eval_episodes"),
"weights": "ema",
"id_envs": required(cfg, "id_envs"),
"ood_envs": required(cfg, "ood_envs"),
},
}
# =============================================================================
# Description
# =============================================================================
def describe(
model_dir: Path, staged: Path, pth: Path, cfg_name: str, stats: dict, cfg: dict,
) -> dict[str, str]:
"""Pull env name and training detail out of a checkpoint's own metadata."""
return {
"path": str(model_dir.relative_to(ROOT)),
"role": ROLES.get(model_dir.parent.name, model_dir.parent.name),
"env": cfg.get("env_name") or ENV_NAME,
"arch": (
f"{cfg['n_layer']}L, d_model {cfg['n_embd']}, {cfg['n_head']} heads, "
f"horizon {cfg['seq_len']}, {stats['params'] / 1e6:.0f}M params"
),
"step": stats["step"],
"detail": stats["detail"],
"size": human_size(staged),
"restores": ", ".join(f"`{k}`" for k in stats["keys"]),
"file": pth.name,
"config": cfg_name,
}
def describe_run(run: Path, staged: Path) -> dict[str, str]:
"""Summarise what an ablation run contributes to the release."""
counts = [
f"{len(list((staged / d).glob('*')))} {d}"
for d in RUN_DIRS if (staged / d).is_dir()
]
files = [f for f in RUN_FILES if (staged / f).is_file()]
return {
"run": run.name,
"path": str(run.relative_to(ROOT)),
"contents": ", ".join([*(f"`{f}`" for f in files), *counts]),
"size": human_size(staged),
}
def describe_inference(name: str, payload: dict) -> dict[str, str]:
"""Summarise one ``--mode inference`` result JSON."""
results = payload.get("results", payload)
per_env = {
env: stats for env, stats in results.items()
if isinstance(stats, dict) and "win_rate" in stats
}
rates = [stats["win_rate"] for stats in per_env.values()]
episodes = {stats.get("n_episodes") for stats in per_env.values()}
return {
"file": name,
"env": (
next(iter(per_env)) if len(per_env) == 1
else f"{len(per_env)} envs" if per_env else "-"
),
"episodes": (
f"{max(e for e in episodes if e)} episodes per env"
if episodes - {None} else "-"
),
"metric": (
f"mean win rate {sum(rates) / len(rates):.2f}" if rates else "-"
),
}
# =============================================================================
# Staging
# =============================================================================
def stage_checkpoints(
staging: Path, models: dict[Path, list[Path]], metric: str | None,
) -> list[dict[str, str]]:
"""Copy each checkpoint directory, scrubbing its provenance metadata."""
rows = []
for model_dir, pths in models.items():
target = staging / model_dir.relative_to(ROOT)
target.mkdir(parents=True, exist_ok=True)
cfg: dict = {}
cfg_name = ""
for src in sorted(model_dir.iterdir()):
if src.suffix in {".yaml", ".yml"}:
cfg, cfg_name = scrub(yaml.safe_load(src.read_text())), src.name
(target / src.name).write_text(yaml.safe_dump(cfg, sort_keys=True))
elif src.suffix == ".pth":
scrub_checkpoint(src, target / src.name)
stats = export_weights(pths[-1], target / "model.safetensors")
(target / "selection.json").write_text(
json.dumps(selection(cfg, stats, metric), indent=2) + "\n",
)
rows.append(describe(model_dir, target, pths[-1], cfg_name, stats, cfg))
return rows
def stage_runs(staging: Path, runs: list[Path]) -> list[dict[str, str]]:
"""Copy each ablation run's summary, tables and figures."""
rows = []
for run in runs:
target = staging / run.relative_to(ROOT)
target.mkdir(parents=True, exist_ok=True)
for name in RUN_FILES:
if (run / name).is_file():
shutil.copy2(run / name, target / name)
for name in RUN_DIRS:
if (run / name).is_dir():
shutil.copytree(run / name, target / name, ignore=COPY_IGNORE)
rows.append(describe_run(run, target))
return rows
def stage_inference(staging: Path, files: list[Path]) -> list[dict[str, str]]:
"""Copy each inference result JSON into ``results/inference/``."""
target_dir = staging / INFERENCE.relative_to(ROOT)
rows: list[dict[str, str]] = []
for src in files:
try:
payload = json.loads(src.read_text())
except json.JSONDecodeError:
print(f"Skipping unreadable inference result {src}.", file=sys.stderr)
continue
name = src.name
if any(r["file"] == name for r in rows):
name = f"{src.parent.name}-{src.name}"
target_dir.mkdir(parents=True, exist_ok=True)
# scrub, not shorten_paths: this shortened cluster paths while doing
# nothing about environment keys, the same asymmetry as stage_runs one
# step further along. These payloads carry no config today.
(target_dir / name).write_text(
json.dumps(scrub(payload), indent=2) + "\n",
)
row = describe_inference(name, payload)
row["size"] = human_size(target_dir / name)
rows.append(row)
return rows
def stage_paper_figures(staging: Path, files: list[Path]) -> list[dict[str, str]]:
"""Copy the manuscript figures into ``results/paper_figures/``."""
if not files:
return []
target_dir = staging / PAPER_FIGURES.relative_to(ROOT)
target_dir.mkdir(parents=True, exist_ok=True)
rows = []
for src in files:
shutil.copy2(src, target_dir / src.name)
rows.append({"file": src.name, "size": human_size(target_dir / src.name)})
return rows
def stage(
staging: Path,
models: dict[Path, list[Path]],
runs: list[Path],
inference: list[Path],
figures: list[Path],
metric: str | None,
) -> tuple[
list[dict[str, str]], list[dict[str, str]],
list[dict[str, str]], list[dict[str, str]],
]:
"""Stage checkpoints, results and LICENSE; the card is written by the caller.
LICENSE comes from git rather than the working tree: a
``hf download --local-dir .`` overwrites it with the Hub's own copy, and
publishing from the tree would push that straight back up as current.
"""
rows = stage_checkpoints(staging, models, metric)
run_rows = stage_runs(staging, runs)
inf_rows = stage_inference(staging, inference)
fig_rows = stage_paper_figures(staging, figures)
# ROOT is passed explicitly rather than read inside the helper, so a
# test can rebind it and have that take effect.
copy_tracked_file("LICENSE", staging / "LICENSE", ROOT)
# Last, so it covers everything every stager copied -- including the whole
# directories stage_runs copies wholesale, which no stager scrubbed.
for name in scrub_staged_tree(staging):
print(f" scrubbed provenance from {name}")
return rows, run_rows, inf_rows, fig_rows
# =============================================================================
# Model card
# =============================================================================
def table(headers: list[str], lines: list[str]) -> str:
sep = "|".join(["---"] * len(headers))
return f"| {' | '.join(headers)} |\n|{sep}|\n" + "".join(f"| {ln} |\n" for ln in lines)
def checkpoint_table(rows: list[dict[str, str]]) -> str:
return table(
["Path", "Role", "Environment", "Architecture", "Selected at", "Training", "Size"],
[
f"`{r['path']}` | {r['role']} | `{r['env']}` | {r['arch']} | "
f"{r['step']} | {r['detail']} | {r['size']}"
for r in sorted(rows, key=lambda r: r["path"])
],
)
def results_section(
run_rows: list[dict[str, str]],
inf_rows: list[dict[str, str]],
fig_rows: list[dict[str, str]],
) -> str:
"""Ablation and inference tables; empty when the release carries neither."""
parts = []
if run_rows:
parts.append(
"RL fine-tuning ablation runs, as produced by "
"`experiments/rl_finetuning/run_ablations.py`. Each run ships its "
"`results.json` summary, the `diagnosis.md` write-up, and the "
"tables (`.csv` and `.tex`) and figures generated from it.\n\n"
+ table(
["Run", "Contents", "Size"],
[
f"`{r['path']}` | {r['contents']} | {r['size']}"
for r in sorted(run_rows, key=lambda r: r["run"])
],
),
)
if inf_rows:
parts.append(
"Evaluation results produced by `main.py --mode inference` on the "
"checkpoints above, under `results/inference/`.\n\n"
+ table(
["File", "Environment", "Evaluation", "Headline metric", "Size"],
[
f"`{r['file']}` | `{r['env']}` | {r['episodes']} | "
f"{r['metric']} | {r['size']}"
for r in sorted(inf_rows, key=lambda r: r["file"])
],
),
)
if fig_rows:
parts.append(
"Manuscript figures as vector PDF, under `results/paper_figures/`. "
"Each puts Craftax Classic and MiniHack side by side, so they are "
"built from both repositories' ablation `results.json` by the "
"sibling repo's `scripts/paper_figures.py` and published "
"identically in both Hub repos.\n\n"
+ table(
["Figure", "Size"],
[
f"`{r['file']}` | {r['size']}"
for r in sorted(fig_rows, key=lambda r: r["file"])
],
),
)
return "## Results\n\n" + "\n".join(parts) if parts else ""
def featured(rows: list[dict[str, str]]) -> dict[str, str]:
"""The checkpoint the download and usage examples are written against."""
planners = [r for r in rows if "planner" in r["role"].lower()]
return sorted(planners or rows, key=lambda r: r["path"])[0]
def model_card(
repo_id: str,
rows: list[dict[str, str]],
run_rows: list[dict[str, str]],
inf_rows: list[dict[str, str]],
fig_rows: list[dict[str, str]],
total_mb: float,
metric: str | None,
) -> str:
example = featured(rows)
selected_on = (
f"selected on {metric}" if metric
else "the metric behind that selection is not recorded in this release"
)
return f"""---
license: mit
library_name: pytorch
pipeline_tag: reinforcement-learning
tags:
- reinforcement-learning
- planning
- discrete-diffusion
- remdm
- minihack
- nethack
- pytorch
---
# ReMDM Planner: {ENV_NAME} checkpoints
Trained weights accompanying *{PAPER}*: a remasking discrete diffusion model
(ReMDM) used as an action-sequence planner in
[MiniHack](https://github.com/facebookresearch/minihack), together with the BFS
oracle rollouts that supervise it, and the results reported in the paper.
Code, configs and evaluation harness: {CODE_URL}
## Contents
{checkpoint_table(rows)}
Each checkpoint ships the `.pth` training state it was published from (weights,
EMA shadow, optimiser, scheduler, and for the DAgger run the curriculum and RNG
state, so training resumes exactly), a `model.safetensors` export of the EMA
weights for inference, the YAML config snapshot it was trained under, and a
`selection.json` recording how it was chosen.
Weights are PyTorch training states with a `safetensors` export of the EMA
weights alongside, and the paths above mirror the source repository so a
snapshot can be dropped straight into a working copy.
{results_section(run_rows, inf_rows, fig_rows)}
## Download
This repo mirrors the code repository's layout, so a snapshot drops straight
into a working copy -- but it also carries its own `README.md` (this card),
`LICENSE` and `.gitattributes`, and `local_dir="."` would overwrite the code
repository's copies of all three. Exclude them, or download into a directory
of its own.
```python
from huggingface_hub import snapshot_download
# everything (~{total_mb:.0f} MB), into a clone of the code repository
snapshot_download(
repo_id="{repo_id}",
local_dir=".",
ignore_patterns=["README.md", "LICENSE", ".gitattributes"],
)
# or somewhere of its own, leaving any working copy untouched
snapshot_download(repo_id="{repo_id}", local_dir="remdm-planner-minihack")
# a single model
snapshot_download(
repo_id="{repo_id}",
local_dir=".",
allow_patterns="{example['path']}/**",
)
```
## Use
From a clone of the code repository, after downloading into it:
```bash
DIR={example['path']}
uv run python main.py --mode inference \\
--config $DIR/{example['config']} --checkpoint $DIR/{example['file']} \\
--output results/inference/eval.json
```
Programmatic loading uses `src.models.denoiser.make_model` with the checkpoint's
own config, then the safetensors export:
```python
from safetensors.torch import load_file
from src.config import load_config
from src.models.denoiser import make_model
cfg = load_config("{example['path']}/{example['config']}")
model = make_model(cfg)
model.load_state_dict(load_file("{example['path']}/model.safetensors"))
model.eval()
```
Architecture arguments should be read from the checkpoint's own config snapshot
rather than from `configs/defaults.yaml`, which tracks the current code.
## Training
The planners are bidirectional transformers that denoise a masked action plan
conditioned on a cropped MiniHack glyph observation, trained either by offline
behaviour cloning on oracle rollouts or by online DAgger against the BFS oracle
under a dynamic environment curriculum. Model size and horizon differ per run
(see the table). Exact hyperparameters for every run, including the
in-distribution and out-of-distribution environment sets, the remasking
strategy, schedule and sampling settings, are in the per-checkpoint config
snapshots listed above, which are the authoritative record.
Both models are best-checkpoint selections rather than final-step dumps: each
trainer evaluates every periodic checkpoint on its configured number of
episodes per environment using EMA weights, and the highest-scoring one is
published ({selected_on}). Directory names encode the sample-equivalents the
published model consumed (gradient steps x batch size, rounded); file names
carry each trainer's own counter, DAgger iterations online and gradient steps
offline. Each checkpoint's `selection.json` records the configured budget it
was drawn from and the step it was selected at.
## Limitations
These are research artefacts tied to specific MiniHack environment versions and
to the cropped-glyph observation encoding; they are not general-purpose agents
and will not transfer to other environments or to pixel observations.
Evaluation results and their variance are reported in the paper.
## Citation
```bibtex
@inproceedings{{remdm-planner-minihack-planner,
title = {{{PAPER}}},
author = {{Anonymous}},
year = {{2026}},
note = {{NeurIPS 2026 Workshop: Beyond Next-Token Prediction}}
}}
```
## License
MIT, see `LICENSE`.
"""
# =============================================================================
# Entry point
# =============================================================================
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
p.add_argument("--repo-id", required=True, help="e.g. <ANON_HF_REPO_ID>")
p.add_argument(
"--inference-results", nargs="*", default=[], metavar="PATH",
help=f"extra --mode inference JSONs or directories, on top of "
f"{INFERENCE.relative_to(ROOT)}/",
)
p.add_argument(
"--selection-metric",
help="metric the best checkpoint was chosen on, e.g. 'mean ID+OOD win rate'",
)
p.add_argument("--private", action="store_true", help="create the repo private")
p.add_argument("--dry-run", action="store_true", help="stage and print, do not upload")
p.add_argument("--yes", action="store_true", help="skip the confirmation prompt")
return p.parse_args()
def main() -> int:
args = parse_args()
token = os.environ.get("HF_TOKEN")
if not args.dry_run and not token:
print("HF_TOKEN is not set.", file=sys.stderr)
return 1
models = discover_checkpoints()
if not models:
print(
f"No .pth checkpoints found under {CKPTS}.\n"
"Discovery expects the released layout, "
"checkpoints/<role>/<name>/*.pth, and skips checkpoints/hf/ "
"because that is where Hub downloads land. Copy a run's "
"checkpoints into checkpoints/{offline,online}/<name> first.",
file=sys.stderr,
)
return 1
runs = discover_runs()
inference = discover_inference(args.inference_results)
figures = discover_paper_figures()
with tempfile.TemporaryDirectory(prefix="remdm-planner-minihack-") as tmp:
staging = Path(tmp)
rows, run_rows, inf_rows, fig_rows = stage(
staging, models, runs, inference, figures, args.selection_metric,
)
total_mb = dir_size_mb(staging)
card = model_card(
args.repo_id, rows, run_rows, inf_rows, fig_rows, total_mb,
args.selection_metric,
)
(staging / "README.md").write_text(card)
files = [f for f in staging.rglob("*") if f.is_file()]
print(f"Staged {plural(len(rows), 'checkpoint')}, "
f"{plural(len(run_rows), 'ablation run')}, "
f"{plural(len(inf_rows), 'inference result')}, "
f"{plural(len(fig_rows), 'paper figure')}, "
f"{plural(len(files), 'file')}, {total_mb:.0f} MB")
for r in sorted(rows, key=lambda r: r["path"]):
print(f" {r['path']:<70} {r['size']:>8} {r['detail']}")
print(f" {'':<70} restores {r['restores']}")
for r in sorted(run_rows, key=lambda r: r["run"]):
print(f" {r['path']:<70} {r['size']:>8} {r['contents']}")
for r in sorted(inf_rows, key=lambda r: r["file"]):
print(f" results/inference/{r['file']:<52} {r['size']:>8} {r['metric']}")
for r in sorted(fig_rows, key=lambda r: r["file"]):
print(f" results/paper_figures/{r['file']:<48} {r['size']:>8}")
if not run_rows:
print(f"Warning: no ablation runs with a results.json under {RUNS}.",
file=sys.stderr)
if not inf_rows:
print("Warning: no inference results; produce them with "
"`main.py --mode inference --output "
f"{INFERENCE.relative_to(ROOT)}/<name>.json`.", file=sys.stderr)
if not fig_rows:
print(
"Warning: no manuscript figures; build them with the sibling "
"repo's `scripts/paper_figures.py` and copy the PDFs into "
f"{PAPER_FIGURES.relative_to(ROOT)}/.", file=sys.stderr,
)
if not args.selection_metric:
print(
"Warning: --selection-metric not given; these are best-of-N "
"checkpoints and the card cannot say what they were chosen on.",
file=sys.stderr,
)
if args.dry_run:
print(f"Dry run; staged tree left nowhere. Card:\n\n{card}")
return 0
if not args.yes:
visibility = "private" if args.private else "public"
if input(f"Upload to {args.repo_id} ({visibility})? [y/N] ").strip().lower() not in {"y", "yes"}:
print("Aborted.")
return 0
from huggingface_hub import HfApi
api = HfApi(token=token)
api.create_repo(args.repo_id, repo_type="model", private=args.private, exist_ok=True)
api.upload_folder(
repo_id=args.repo_id,
folder_path=str(staging),
repo_type="model",
ignore_patterns=HUB_IGNORE,
commit_message="Upload MiniHack ReMDM planner checkpoints and results",
)
print(f"Done: https://huggingface.co/{args.repo_id}")
return 0
if __name__ == "__main__":
sys.exit(main())
|