Place training config bundles alongside each model and simplify documentation

README.md

@@ -45,7 +45,7 @@ discovery for all models in the table. Install it in a Python environment:
 ```bash
 python -m pip install 'huggingface_hub>=1.0,<2'
 hf auth login
-hf download maltefranke/morph --revision da1c241d83f2628b172fc510e1bc724ef4429db2 \
+hf download maltefranke/morph \
   --local-dir morph-runtime \
   --include 'geom/explicit_h/runtime/*'
 python -m pip install ./morph-runtime/geom/explicit_h/runtime
@@ -138,83 +138,28 @@ supports all three published model variants.
 
 ## Reproducibility
 
-### Generate with a fixed release
-
-Pin the **runtime code and dependencies** as well as the **model bundle**.
-This installs the archived shared inference runtime, independently of changes
-to this repository or the model catalog. Install [uv](https://docs.astral.sh/uv/)
-and authenticate with an account that can access the model repository:
-
-```bash
-uv tool install 'huggingface_hub==1.33.0'
-hf auth login
-hf download maltefranke/morph \
-  --revision da1c241d83f2628b172fc510e1bc724ef4429db2 \
-  --include 'geom/explicit_h/runtime/*' --local-dir morph-runtime
-cd morph-runtime/geom/explicit_h/runtime
-uv sync --frozen --no-dev --python 3.13.7
-source .venv/bin/activate
-
-# Run on a compute worker. Save all attempts and their validity flags.
-MORPH_COMPILE=0 MORPH_ATTENTION_BACKEND=sdpa MORPH_FLASH_ATTN=0 \
-morph generate --model maltefranke/morph --subfolder geom/explicit_h \
-  --revision b839af108eaade757a5d2d668d05498e8c733c1e --epoch 50 \
-  --device cuda --ema --steps 100 --seed 42 --batch 64 --batch-size 16 \
-  --save-trajs --output geom-e50-seed42.sdf
-uv pip freeze > environment.txt
-nvidia-smi --query-gpu=name,driver_version --format=csv > gpu.csv
-```
-
-Choose another published run by changing these three options:
-
-| `--subfolder` | `--revision` | `--epoch` |
-| --- | --- | --- |
-| `chembl3d/explicit_h` | `ab9bea92f499f03450ebc99459fce04461504a41` | `200`, `300`, `400` |
-| `chembl3d/implicit_h` | `ab9bea92f499f03450ebc99459fce04461504a41` | `100`, `200`, `300` |
-| `geom/explicit_h` | `b839af108eaade757a5d2d668d05498e8c733c1e` | `50` |
-
-Keep the seed, batch size, integration steps, EMA selection, precision, attention
-backend and software/hardware environment fixed when comparing samples. A seed
-alone does not guarantee identical samples across GPU models or library versions.
-The checkpoint SHA-256 hashes are in each bundle's `metadata.json`.
+Each model has a **training bundle** with its original training configuration,
+launch settings, frozen source, dependency versions and data references.
+The bundles live alongside the corresponding model files on Hugging Face:
 
-### Repeat a training recipe
+| Model | Configuration and source |
+| --- | --- |
+| ChEMBL3D, explicit H | [Training bundle](https://huggingface.co/maltefranke/morph/resolve/main/chembl3d/explicit_h/training_bundle.tar.gz) |
+| ChEMBL3D, implicit H | [Training bundle](https://huggingface.co/maltefranke/morph/resolve/main/chembl3d/implicit_h/training_bundle.tar.gz) |
+| GEOM-Drugs, explicit H | [Training bundle](https://huggingface.co/maltefranke/morph/resolve/main/geom/explicit_h/training_bundle.tar.gz) |
+| QM9, explicit H (no published weights yet) | [Training bundle](https://huggingface.co/maltefranke/morph/resolve/main/qm9/explicit_h/training_bundle.tar.gz) |
 
-The [training archives](https://huggingface.co/maltefranke/morph/tree/main/reproduction)
-contain the original source (including `run.py`), saved training config, launcher,
-dependency versions, and a portable `train.sh`. They use **2 RTX 4090s, seed 42,
-and bf16 mixed precision**. Run installation and training in a compute allocation:
+For example, download and unpack the GEOM training bundle:
 
 ```bash
-# Start from a directory outside the inference environment above.
-run=geom-explicit-h  # Or chembl3d-explicit-h, chembl3d-implicit-h, qm9-explicit-h
-hf download maltefranke/morph --revision d0332f1c022345956470bb948ada79e5702ed8dc \
-  "reproduction/$run.tar.gz" reproduction/SHA256SUMS --local-dir morph-training
-cd morph-training/reproduction
-sha256sum --ignore-missing -c SHA256SUMS
-tar -xzf "$run.tar.gz"
-cd "$run"
-uv venv --python 3.13.7 .venv
-uv pip install --python .venv/bin/python -r pyproject.toml \
-  -c reproduction/requirements-environment.txt
-source .venv/bin/activate
-export MORPH_DATASET_ROOT=/absolute/path/to/geom  # Contains processed/
-export MORPH_OUTPUT_DIR="$PWD/outputs/repeat-01"
-bash train.sh
+hf download maltefranke/morph geom/explicit_h/training_bundle.tar.gz --local-dir .
+tar -xzf geom/explicit_h/training_bundle.tar.gz
 ```
 
-All portable recipes evaluate every epoch and retain every fifth epoch; QM9
-retains every **25th** epoch. The archives start fresh and stop at 400 epochs for
-explicit-H ChEMBL3D, 300 for implicit-H ChEMBL3D, 50 for GEOM, or 3000 for QM9.
-Override the stopping point with `bash train.sh trainer.trainer.max_epochs=100`.
-QM9 currently has a training recipe but no published checkpoint.
-
-These commands reproduce the documented setup, not an exact replay of every
-historical run: explicit-H ChEMBL3D resumed at epoch 5 and GEOM at epoch 48;
-the older explicit-H ChEMBL3D run also evaluated every five epochs. Data caches
-and those intermediate resume checkpoints are not included in the archives.
-See [reproducibility details](reproduction/README.md) for the data requirements,
-original settings, environment provenance and resume instructions.
+Each bundle also includes a portable training recipe; its changes from the
+original run are recorded alongside it. See the [training bundle guide](reproduction/README.md)
+for setup, data requirements and resuming. Version identifiers and checksums
+are recorded in the release metadata and bundle manifests.
 
 ## Hydrogen handling and training data
 
@@ -253,16 +198,17 @@ Each model directory contains:
   filenames and checkpoint SHA-256 hashes.
 - `training_config.yaml`: the original training configuration for provenance.
 - `runtime/`: archived source, configs and dependency specifications.
+- `training_bundle.tar.gz`: original run configuration, source, environment and a
+  portable training recipe; `training_bundle.json` records its checksum and provenance.
 - `validation.json`: results of checkpoint loading and sampling checks.
 
 The published bundles were checked for strict current/EMA weight loading and
 end-to-end sampling. These are functionality checks, not validity benchmarks.
 The GEOM release also passed remote file-size and hash verification.
 
-The [ChEMBL3D release manifest](https://huggingface.co/maltefranke/morph/blob/ab9bea92f499f03450ebc99459fce04461504a41/manifest.json)
-describes the original ChEMBL3D release at its pinned revision, including its
-historical model card. GEOM has its own
-[bundle manifest](https://huggingface.co/maltefranke/morph/blob/b839af108eaade757a5d2d668d05498e8c733c1e/geom/explicit_h/manifest.json).
+The [ChEMBL3D release manifest](https://huggingface.co/maltefranke/morph/blob/main/manifest.json)
+describes the original ChEMBL3D release, including its historical model card. GEOM has its own
+[bundle manifest](https://huggingface.co/maltefranke/morph/blob/main/geom/explicit_h/manifest.json).
 Original training configs contain historical cluster paths; use `config.yaml`
 through the inference loader for portable generation.
 

chembl3d/explicit_h/training_bundle.json

@@ -0,0 +1,57 @@
+{
+  "model": "chembl3d-explicit-h",
+  "source_snapshot": "chembl3d_recovery_20260927_v2",
+  "published_completed_epochs": [
+    200,
+    300,
+    400
+  ],
+  "original_resume_completed_epoch": 5,
+  "original_validation_every_epochs": 5,
+  "reproduction_validation_every_epochs": 1,
+  "checkpoint_every_epochs": 5,
+  "seed": 42,
+  "devices": 2,
+  "gpu": "NVIDIA GeForce RTX 4090",
+  "batch_per_gpu": 320,
+  "accumulation": 1,
+  "precision": "bf16-mixed",
+  "recipe_changes": [
+    "Portable data/output paths",
+    "Fresh initialization; no historical resume",
+    "Epoch limit instead of allocation timer",
+    "W&B disabled",
+    "Evaluation every epoch; best and retained checkpoints every 5 epochs (QM9: 25)"
+  ],
+  "environment_note": "Python 3.13.7. Installed package versions captured from the shared training environment on 2026-10-08; not an original per-job environment lock.",
+  "limits": "Data and historical resume checkpoints are not distributed. Same recipe and seed do not imply a bitwise replay across resumes, hardware or software changes.",
+  "archive": "training_bundle.tar.gz",
+  "bytes": 422898,
+  "sha256": "d72608eb7ce5ce3fced959544c7c5a869ff6ae574d2b9e87e2d7aa1ac0ac1f4a",
+  "directory": "chembl3d-explicit-h",
+  "model_folder": "chembl3d/explicit_h",
+  "source_release": {
+    "repo_id": "maltefranke/morph",
+    "revision": "d0332f1c022345956470bb948ada79e5702ed8dc",
+    "filename": "reproduction/chembl3d-explicit-h.tar.gz"
+  },
+  "checkpoint_release": {
+    "repo_id": "maltefranke/morph",
+    "subfolder": "chembl3d/explicit_h",
+    "revision": "ab9bea92f499f03450ebc99459fce04461504a41",
+    "epochs": [
+      200,
+      300,
+      400
+    ]
+  },
+  "files": {
+    "original_config": "reproduction/original_training_config.yaml",
+    "portable_config": "configs/reproduce.yaml",
+    "original_launcher": "reproduction/original_launcher.sbatch",
+    "environment": "reproduction/environment.sh",
+    "dependencies": "reproduction/requirements-environment.txt",
+    "data_reference": "reproduction/dataset_reference.json",
+    "provenance": "reproduction/provenance.json"
+  }
+}

chembl3d/explicit_h/training_bundle.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d72608eb7ce5ce3fced959544c7c5a869ff6ae574d2b9e87e2d7aa1ac0ac1f4a
+size 422898

chembl3d/implicit_h/training_bundle.json

@@ -0,0 +1,57 @@
+{
+  "model": "chembl3d-implicit-h",
+  "source_snapshot": "chembl3d_implicit_20260928_v2",
+  "published_completed_epochs": [
+    100,
+    200,
+    300
+  ],
+  "original_resume_completed_epoch": null,
+  "original_validation_every_epochs": 1,
+  "reproduction_validation_every_epochs": 1,
+  "checkpoint_every_epochs": 5,
+  "seed": 42,
+  "devices": 2,
+  "gpu": "NVIDIA GeForce RTX 4090",
+  "batch_per_gpu": 320,
+  "accumulation": 1,
+  "precision": "bf16-mixed",
+  "recipe_changes": [
+    "Portable data/output paths",
+    "Fresh initialization; no historical resume",
+    "Epoch limit instead of allocation timer",
+    "W&B disabled",
+    "Evaluation every epoch; best and retained checkpoints every 5 epochs (QM9: 25)"
+  ],
+  "environment_note": "Python 3.13.7. Installed package versions captured from the shared training environment on 2026-10-08; not an original per-job environment lock.",
+  "limits": "Data and historical resume checkpoints are not distributed. Same recipe and seed do not imply a bitwise replay across resumes, hardware or software changes.",
+  "archive": "training_bundle.tar.gz",
+  "bytes": 241231,
+  "sha256": "7add66a5db756b459e15c6c9d0bf75bf6110d007c17f2cc30fbcde4c58ef9695",
+  "directory": "chembl3d-implicit-h",
+  "model_folder": "chembl3d/implicit_h",
+  "source_release": {
+    "repo_id": "maltefranke/morph",
+    "revision": "d0332f1c022345956470bb948ada79e5702ed8dc",
+    "filename": "reproduction/chembl3d-implicit-h.tar.gz"
+  },
+  "checkpoint_release": {
+    "repo_id": "maltefranke/morph",
+    "subfolder": "chembl3d/implicit_h",
+    "revision": "ab9bea92f499f03450ebc99459fce04461504a41",
+    "epochs": [
+      100,
+      200,
+      300
+    ]
+  },
+  "files": {
+    "original_config": "reproduction/original_training_config.yaml",
+    "portable_config": "configs/reproduce.yaml",
+    "original_launcher": "reproduction/original_launcher.sbatch",
+    "environment": "reproduction/environment.sh",
+    "dependencies": "reproduction/requirements-environment.txt",
+    "data_reference": "reproduction/dataset_reference.json",
+    "provenance": "reproduction/provenance.json"
+  }
+}

chembl3d/implicit_h/training_bundle.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7add66a5db756b459e15c6c9d0bf75bf6110d007c17f2cc30fbcde4c58ef9695
+size 241231

geom/explicit_h/training_bundle.json

@@ -0,0 +1,53 @@
+{
+  "model": "geom-explicit-h",
+  "source_snapshot": "geom_resume_e48_20261008",
+  "published_completed_epochs": [
+    50
+  ],
+  "original_resume_completed_epoch": 48,
+  "original_validation_every_epochs": 1,
+  "reproduction_validation_every_epochs": 1,
+  "checkpoint_every_epochs": 5,
+  "seed": 42,
+  "devices": 2,
+  "gpu": "NVIDIA GeForce RTX 4090",
+  "batch_per_gpu": 320,
+  "accumulation": 1,
+  "precision": "bf16-mixed",
+  "recipe_changes": [
+    "Portable data/output paths",
+    "Fresh initialization; no historical resume",
+    "Epoch limit instead of allocation timer",
+    "W&B disabled",
+    "Evaluation every epoch; best and retained checkpoints every 5 epochs (QM9: 25)"
+  ],
+  "environment_note": "Python 3.13.7. Installed package versions captured from the shared training environment on 2026-10-08; not an original per-job environment lock.",
+  "limits": "Data and historical resume checkpoints are not distributed. Same recipe and seed do not imply a bitwise replay across resumes, hardware or software changes.",
+  "archive": "training_bundle.tar.gz",
+  "bytes": 423904,
+  "sha256": "d61d4edc6c530d8efe03016316beecdb30cff1b8228f7c433bb4f98f59606695",
+  "directory": "geom-explicit-h",
+  "model_folder": "geom/explicit_h",
+  "source_release": {
+    "repo_id": "maltefranke/morph",
+    "revision": "d0332f1c022345956470bb948ada79e5702ed8dc",
+    "filename": "reproduction/geom-explicit-h.tar.gz"
+  },
+  "checkpoint_release": {
+    "repo_id": "maltefranke/morph",
+    "subfolder": "geom/explicit_h",
+    "revision": "b839af108eaade757a5d2d668d05498e8c733c1e",
+    "epochs": [
+      50
+    ]
+  },
+  "files": {
+    "original_config": "reproduction/original_training_config.yaml",
+    "portable_config": "configs/reproduce.yaml",
+    "original_launcher": "reproduction/original_launcher.sbatch",
+    "environment": "reproduction/environment.sh",
+    "dependencies": "reproduction/requirements-environment.txt",
+    "data_reference": "reproduction/dataset_reference.json",
+    "provenance": "reproduction/provenance.json"
+  }
+}

geom/explicit_h/training_bundle.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d61d4edc6c530d8efe03016316beecdb30cff1b8228f7c433bb4f98f59606695
+size 423904

qm9/explicit_h/training_bundle.json

@@ -0,0 +1,44 @@
+{
+  "model": "qm9-explicit-h",
+  "source_snapshot": "qm9_launch_20261008",
+  "published_completed_epochs": [],
+  "original_resume_completed_epoch": null,
+  "original_validation_every_epochs": 1,
+  "reproduction_validation_every_epochs": 1,
+  "checkpoint_every_epochs": 25,
+  "seed": 42,
+  "devices": 2,
+  "gpu": "NVIDIA GeForce RTX 4090",
+  "batch_per_gpu": 1024,
+  "accumulation": 2,
+  "precision": "bf16-mixed",
+  "recipe_changes": [
+    "Portable data/output paths",
+    "Fresh initialization; no historical resume",
+    "Epoch limit instead of allocation timer",
+    "W&B disabled",
+    "Evaluation every epoch; best and retained checkpoints every 5 epochs (QM9: 25)"
+  ],
+  "environment_note": "Python 3.13.7. Installed package versions captured from the shared training environment on 2026-10-08; not an original per-job environment lock.",
+  "limits": "Data and historical resume checkpoints are not distributed. Same recipe and seed do not imply a bitwise replay across resumes, hardware or software changes.",
+  "archive": "training_bundle.tar.gz",
+  "bytes": 423543,
+  "sha256": "06aa71e19be7423097ccb08540a463bb7f4fa6d99152428fff0cc47eb0ae8c8f",
+  "directory": "qm9-explicit-h",
+  "model_folder": "qm9/explicit_h",
+  "source_release": {
+    "repo_id": "maltefranke/morph",
+    "revision": "d0332f1c022345956470bb948ada79e5702ed8dc",
+    "filename": "reproduction/qm9-explicit-h.tar.gz"
+  },
+  "checkpoint_release": null,
+  "files": {
+    "original_config": "reproduction/original_training_config.yaml",
+    "portable_config": "configs/reproduce.yaml",
+    "original_launcher": "reproduction/original_launcher.sbatch",
+    "environment": "reproduction/environment.sh",
+    "dependencies": "reproduction/requirements-environment.txt",
+    "data_reference": "reproduction/dataset_reference.json",
+    "provenance": "reproduction/provenance.json"
+  }
+}

qm9/explicit_h/training_bundle.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06aa71e19be7423097ccb08540a463bb7f4fa6d99152428fff0cc47eb0ae8c8f
+size 423543

reproduction/README.md

@@ -1,76 +1,61 @@
-# Reproducing models and training runs
+# Training bundles
 
-The [README](../README.md#reproducibility) contains complete commands for pinned
-inference and training. The model repository is
-[maltefranke/morph](https://huggingface.co/maltefranke/morph).
+Each model folder in [maltefranke/morph](https://huggingface.co/maltefranke/morph)
+contains a `training_bundle.tar.gz` with the original configuration, source and
+run settings. The [README](../README.md#reproducibility) links to each bundle.
+No version suffix or commit identifier is needed in the commands below.
 
-## Published inference
+## Download and set up
 
-Pin both the installed source and the checkpoint/assets. The shared inference
-runtime at Hub revision `da1c241d83f2628b172fc510e1bc724ef4429db2` supports all
-published models; its `uv.lock` pins dependencies. The README installs that
-runtime with Python 3.13.7. Installing a different checkout of Morph changes the
-implementation even when `load_pretrained(..., revision=...)` selects the same
-weights: that argument pins the model bundle, not the installed Python code.
+With [uv](https://docs.astral.sh/uv/) installed and `hf auth login` completed,
+choose a model folder and download its bundle:
 
-With that environment active, this is the Python equivalent of the pinned
-GEOM example. Set the environment variables before importing Morph:
+```bash
+hf download maltefranke/morph geom/explicit_h/training_bundle.tar.gz \
+  geom/explicit_h/training_bundle.json --local-dir morph-training
+cd morph-training
+```
+
+The small JSON file records the archive checksum, original run settings and
+checkpoint release. To verify the download before unpacking:
 
 ```python
-import os
-
-os.environ["MORPH_COMPILE"] = "0"
-os.environ["MORPH_ATTENTION_BACKEND"] = "sdpa"
-os.environ["MORPH_FLASH_ATTN"] = "0"
-
-from morph.inference import load_pretrained
-
-model = load_pretrained(
-    "maltefranke/morph",
-    subfolder="geom/explicit_h",
-    revision="b839af108eaade757a5d2d668d05498e8c733c1e",
-    epoch=50,
-    device="cuda",
-    steps=100,
-    use_ema=True,
-)
-trajectories, validity, vocab, coordinate_std = model.generate(
-    n_mols=64, batch_size=16, seed=42, return_traj=True,
-)
-print("Model revision:", model.bundle.revision)
-print("Valid attempts:", sum(validity), "/", len(validity))
+import hashlib
+import json
+from pathlib import Path
+
+folder = Path("geom/explicit_h")
+metadata = json.loads((folder / "training_bundle.json").read_text())
+with (folder / metadata["archive"]).open("rb") as stream:
+    assert hashlib.file_digest(stream, "sha256").hexdigest() == metadata["sha256"]
 ```
 
-Direct repository IDs bypass catalog lookup. To use a short model name while
-freezing discovery, pass a pinned catalog:
+Install and train inside a compute allocation:
 
-```python
-from morph.inference import load_catalog, load_pretrained
-
-catalog = load_catalog(
-    repo_id="maltefranke/morph",
-    revision="377e8b52505d564df8529fe6fef2da9c77f57661",
-)
-model = load_pretrained(
-    "chembl3d-implicit-h", catalog=catalog, epoch=300,
-    device="cuda", steps=100, use_ema=True,
-)
+```bash
+tar -xzf geom/explicit_h/training_bundle.tar.gz
+cd geom-explicit-h
+uv venv --python 3.13.7 .venv
+uv pip install --python .venv/bin/python -r pyproject.toml \
+  -c reproduction/requirements-environment.txt
+source .venv/bin/activate
+export MORPH_DATASET_ROOT=/absolute/path/to/geom  # Contains processed/
+export MORPH_OUTPUT_DIR="$PWD/outputs/repeat-01"
+bash train.sh
 ```
 
-Changing batch size can change random-number consumption and generated samples.
-Record the generation count, batch size, seed, steps, EMA/current choice,
-precision, backend, source revision, package versions, GPU and driver. The CLI's
-SDF includes valid molecules only; `--save-trajs` preserves every attempt and
-its validity flag. Coordinates are already in Angstrom.
+For another model, use its folder from the table below. The extracted directory
+is the model name, also recorded as `directory` in `training_bundle.json`.
 
-Seeds and fixed releases provide controlled comparisons, but CUDA kernels and
-changes to hardware, dependencies or precision can prevent bitwise equality.
-The historical training configuration uses `deterministic: warn`, which is not
-a guarantee that every operation has a deterministic implementation.
+| Model folder | Extracted directory |
+| --- | --- |
+| `chembl3d/explicit_h` | `chembl3d-explicit-h` |
+| `chembl3d/implicit_h` | `chembl3d-implicit-h` |
+| `geom/explicit_h` | `geom-explicit-h` |
+| `qm9/explicit_h` | `qm9-explicit-h` |
 
-## Training archives
+## Bundle contents
 
-Download from `reproduction/` at immutable Hub revision `d0332f1c022345956470bb948ada79e5702ed8dc`.
 Each archive contains:
 
 - The original `run.py`, `src/`, `configs/`, `pyproject.toml`, and `uv.lock`
@@ -88,20 +73,22 @@ Each archive contains:
   per-job environment capture; Python 3.13.7 is the reference interpreter.
 
 The archives are roughly 0.2–0.5 MB each and contain no datasets or weights.
-`reproduction/SHA256SUMS` verifies the archives; each archive's manifest verifies
-its files. Preserve the archive and immutable revision with your results.
+The adjacent `training_bundle.json` verifies the archive; each archive's manifest
+verifies its files. The metadata records the underlying source release and model
+release, so their identifiers stay with the bundle. Preserve the downloaded
+archive and metadata with your results.
 
-The new launcher records package versions, GPU/driver information, requested
+The launcher records package versions, GPU/driver information, requested
 overrides, recipe provenance and data references in `MORPH_OUTPUT_DIR`.
 Hydra records the effective configuration in `.hydra/`. Choose a new output
 directory for each run, and retain these records alongside checkpoints.
 
-| Archive | Batch per GPU | Accumulation | Effective batch | Original validation cadence | Portable validation / save cadence |
+| Model | Batch per GPU | Accumulation | Effective batch | Original validation cadence | Portable validation / save cadence |
 | --- | ---: | ---: | ---: | --- | --- |
-| `chembl3d-explicit-h.tar.gz` | 320 | 1 | 640 | Every 5 epochs | Every 1 / 5 epochs |
-| `chembl3d-implicit-h.tar.gz` | 320 | 1 | 640 | Every epoch | Every 1 / 5 epochs |
-| `geom-explicit-h.tar.gz` | 320 | 1 | 640 | Every epoch | Every 1 / 5 epochs |
-| `qm9-explicit-h.tar.gz` | 1024 | 2 | 4096 | Every epoch | Every 1 / 25 epochs |
+| `chembl3d-explicit-h` | 320 | 1 | 640 | Every 5 epochs | Every 1 / 5 epochs |
+| `chembl3d-implicit-h` | 320 | 1 | 640 | Every epoch | Every 1 / 5 epochs |
+| `geom-explicit-h` | 320 | 1 | 640 | Every epoch | Every 1 / 5 epochs |
+| `qm9-explicit-h` | 1024 | 2 | 4096 | Every epoch | Every 1 / 25 epochs |
 
 All recipes use two RTX 4090 GPUs, seed 42, bf16 mixed precision, SDPA attention,
 compiled networks and 4096 generation attempts per validation epoch. GEOM uses
@@ -109,6 +96,10 @@ Muon with CUDA graph execution; ChEMBL3D and QM9 use native Muon execution with
 DDP compiler graph splitting disabled and compilation fallback enabled.
 Environment flags are explicit in each archive's `reproduction/environment.sh`.
 
+The portable recipes stop at 400 completed epochs for explicit-H ChEMBL3D,
+300 for implicit-H ChEMBL3D, 50 for GEOM, or 3000 for QM9. Override that limit
+with `bash train.sh trainer.trainer.max_epochs=100`.
+
 The portable configs preserve the saved model, optimizer, schedule and batch
 settings. They change data/output paths, disable W&B, remove the historical
 resume path and allocation timer, and use a finite epoch limit. They also apply