Publish compact three-model CD-LAM release

.gitattributes

@@ -1,35 +1,4 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
 *.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text

CITATION.cff

@@ -1,8 +1,7 @@
 cff-version: 1.2.0
-message: "If you use CD-LAM, please cite the manuscript and this software release."
+message: "If you use CD-LAM, please cite the manuscript and software release."
 title: "Causally Debiased Latent Action Model for Embodied Action Conditioned World Models"
 type: software
-license: Apache-2.0
 repository-code: "https://github.com/yufanwei/CD-LAM"
 repository-artifact: "https://huggingface.co/yufanwei/CD-LAM"
 url: "https://yufanwei.github.io/CD-LAM-project-page/"
@@ -34,7 +33,6 @@ authors:
 preferred-citation:
   type: article
   title: "Causally Debiased Latent Action Model for Embodied Action Conditioned World Models"
-  year: 2026
   authors:
     - given-names: Yufan
       family-names: Wei

README.md

@@ -11,99 +11,156 @@ tags:
   - research
 ---
 
-# CD-LAM 2B checkpoints
+# CD-LAM checkpoints
 
-This repository contains the upload-ready 2B checkpoint family for CD-LAM
-(Causally Debiased Latent Action Model). It provides two internally compatible
-lineages: the main 100h lineage and a separate 1000h data-scaling lineage.
+This repository publishes three main model entries for CD-LAM (Causally
+Debiased Latent Action Model): a latent-action model, a compatible 2B
+pretraining overlay, and the selected 100h posttraining overlay.
 
-Each checkpoint was created from a hash-gated research snapshot by retaining
-only the tensors and small compatibility metadata needed at inference time.
-Every released file was reloaded and compared tensor-for-tensor with its
-selected source state. The adjacent JSON sidecar records the source hash,
-release hash, tensor schema, optimizer step, model scale, data tier, and
-compatibility identifiers. `asset_manifest.json` is the aggregate release
-index. `verification_report.json` records the upload-safety audit, and
-`SHA256SUMS` covers every other file in this repository.
+Each checkpoint is produced from a hash-gated research snapshot by retaining
+only the tensors and compatibility metadata required at inference time. Every
+released tensor is reloaded and compared with its selected source state.
+`asset_manifest.json` is the authoritative index for formats, byte sizes,
+SHA-256 values, optimizer steps, model roles, the public runtime commit, and
+compatibility identifiers. License and access terms are summarized separately
+in `MODEL_LICENSE_NOTICE.md` and `external_dependencies.json`.
 
-After downloading a complete snapshot, verify it with:
+## Download and verify
+
+Download one immutable snapshot rather than resolving individual files at
+runtime:
 
 ```bash
+hf download yufanwei/CD-LAM \
+  --revision <immutable-HF-commit> \
+  --local-dir checkpoints/CD-LAM
+cd checkpoints/CD-LAM
 sha256sum -c SHA256SUMS
 ```
 
-## Released assets
+The CD-LAM source repository then consumes only local paths. Keep the snapshot
+revision fixed in reproducible environments.
+
+## Main model entries
 
-| lineage | component | path | status |
+| entry | path | checkpoint role | status |
 |---|---|---|---|
-| 100h | Stage 1 LAM, step 149 | `models/lam/100h/stage1-step149/model.pt` | research; canonical downstream latent space |
-| 100h | Stage 2 ACWM overlay, step 2000 | `models/2b/100h/stage2-step2000/model.pt` | unverified paper role; Stage 3 parent |
-| 100h | 22D-to-32D bridge, epoch 300 | `models/bridge/100h/epoch300/model.pt` | research; canonical downstream bridge |
-| 100h | Stage 3 ACWM overlay, step 3000 | `models/2b/100h/stage3-step3000/model.pt` | unverified paper-budget candidate |
-| 1000h | Stage 1 LAM, step 180 | `models/lam/1000h/stage1-step180/model.pt` | research; canonical downstream latent space |
-| 1000h | Stage 2 ACWM overlay, step 2000 | `models/2b/1000h/stage2-step2000/model.pt` | unverified research lineage |
-| 1000h | 22D-to-32D bridge, epoch 300 | `models/bridge/1000h/epoch300/model.pt` | research; canonical downstream bridge |
-| 1000h | Stage 3 ACWM overlay, step 3000 | `models/2b/1000h/stage3-step3000/model.pt` | research/data-scaling lineage |
-
-The 100h Stage 3 file uses the manuscript's 3,000-update budget and its
-available PSNR matches the manuscript, but the exact FDCE attribution remains
-unresolved. It must not be presented as a fully verified reproduction of the
-paper table. The 1000h lineage is a research scaling result, not a manuscript
-headline checkpoint.
-
-Do not mix assets across the 100h and 1000h rows. Each Stage 2, bridge, and
-Stage 3 sidecar records the SHA-256 of its matching LAM latent space. The 1000h
-step-180 LAM is intentional; a later research checkpoint is not a compatible
-replacement.
+| LAM | `models/lam/model.pt` | masked 32D latent-action encoder, step 300 | research |
+| Pretrain | `models/pretrain/model.pt` | 2B world-model pretraining overlay, step 4000 | research; compatible with the LAM entry |
+| Posttrain (100h) | `models/posttrain-100h/model.pt` | 2B robot-action posttraining overlay, step 3000 | unverified main-result checkpoint |
+
+Internal experiment-version labels are intentionally absent from the public
+paths. Use the semantic entry names above and the hashes in
+`asset_manifest.json`.
+
+## Compatibility is explicit
+
+These three entries are not one direct training chain.
+
+The LAM and Pretrain entries form a compatible baseline pair. The 100h
+posttraining entry belongs to a separate latent-space lineage: it was trained
+with a 100h LAM checkpoint at step 149 and its matching action-to-latent
+bridge. Do not encode a latent with `models/lam/model.pt` and pass it to
+`models/posttrain-100h/model.pt` as if the public LAM were the posttraining
+model's encoder.
+
+For robot-action inference, use the bridge and action contract colocated with
+the posttraining model:
+
+| auxiliary file | path | purpose |
+|---|---|---|
+| action-to-latent bridge | `models/posttrain-100h/bridge.pt` | maps the documented 22D AgiBot action delta into the posttraining model's 32D latent space |
+| action contract | `models/posttrain-100h/action_contract.json` | fixes component order, stride, representation, and compatibility identifiers |
+
+The bridge is part of the Posttrain entry; it is not a fourth main model.
+Missing bridge normalization statistics or a missing action contract make the
+posttraining package incomplete.
+
+Both world-model entries are overlays rather than redistributed base models.
+They require the compatible external NVIDIA 2B base checkpoint, tokenizer,
+text encoder, and the public CD-LAM runtime overlay applied to the pinned base
+source. `asset_manifest.json` binds the exact CD-LAM runtime commit, NVIDIA
+source commit, base-model repository revision, base-model path, file count,
+byte count, and tree-manifest digest. Do not silently substitute another base
+checkpoint or a similarly named LAM.
+
+The required base model is
+`nvidia/DreamDojo@89d029e10816d2995d700cb8ba06f171e0504203`, path
+`2B_pretrain/iter_000140000/model`. It contains 257 files totaling
+12,925,543,625 bytes. Download it under the NVIDIA model terms and keep the
+revision immutable.
 
 ## Checkpoint formats
 
-The files are plain PyTorch checkpoints:
+The files are ordinary PyTorch checkpoints:
 
 ```python
 import torch
 
-checkpoint = torch.load("model.pt", map_location="cpu", weights_only=True)
+checkpoint = torch.load("models/lam/model.pt", map_location="cpu", weights_only=True)
 print(checkpoint["format"])
 print(checkpoint["metadata"])
 ```
 
-- Stage 1 uses format `cdlam.stage1.inference` and contains `model`, plus the
-  released calibration/head state when present.
-- Stages 2 and 3 use `cdlam.stage2.overlay` and `cdlam.stage3.overlay`. They
-  contain `trainable_state`, which must be applied to the matching external 2B
-  base model by the source repository's production adapter.
-- Bridge files use `cdlam.action_to_latent_bridge`. They contain `g_state`, the
-  optional diagnostic inverse `h_state`, and the complete `action_mean`,
-  `action_std`, `zm`, and `zsd` normalization contract.
+- LAM uses `cdlam.stage1.inference` and contains `model`, plus released
+  calibration or diagnostic head state when present.
+- Pretrain uses `cdlam.stage2.overlay` and contains `trainable_state`.
+- Posttrain uses `cdlam.stage3.overlay` and contains `trainable_state`.
+- The bridge uses `cdlam.action_to_latent_bridge` and contains `g_state`, the
+  optional diagnostic inverse `h_state`, and complete `action_mean`,
+  `action_std`, `zm`, and `zsd` normalization tensors.
 
-The base model, tokenizer, text encoder, datasets, SAM3, and CoWTracker are not
-redistributed here. See `external_dependencies.json` and the
-[CD-LAM source repository](https://github.com/yufanwei/CD-LAM) for pinned
-dependency and runtime instructions.
-
-Built on NVIDIA Cosmos
+Loaders should verify the local file SHA-256 and metadata before applying any
+state. Overlay keys must match the expected base architecture; partial or
+unexpected-key loading is an error.
 
 ## Action contract
 
-Stage 3 and bridge assets use contract
-`agibot-alpha-22d-raw-adjacent-stride4`. The input is a raw adjacent
-stride-4 delta in this fixed order: left arm 7, right arm 7, left gripper 1,
-right gripper 1, head 2, waist 2, and base velocity 2. The full machine-readable
-contract is in `action_contract.json`.
+The posttraining bridge uses contract
+`agibot-alpha-22d-raw-adjacent-stride4`. Its input is a raw adjacent stride-4
+delta ordered as left arm 7, right arm 7, left gripper 1, right gripper 1,
+head 2, waist 2, and base velocity 2.
+
+The bridge computes:
+
+```text
+u_norm = (u - action_mean) / action_std
+z_norm = g_state(u_norm)
+z      = z_norm * zsd + zm
+```
 
-Normalization is part of the bridge checkpoint. Do not substitute dataset
-statistics from another lineage, change absolute actions to deltas without the
-documented conversion, or feed packed action chunks directly to the 22D
-bridge.
+Do not replace deltas with absolute commands, change the cadence, reorder
+components, or substitute statistics from another dataset. Shape checks alone
+cannot detect those errors. The contract also pins the AgiBot statistics and
+modality JSON files to the public NVIDIA source revision recorded in the
+manifest, including their relative paths and SHA-256 values.
+
+## Verification boundary
+
+This is an inference and evaluation artifact distribution, not a
+training-resume snapshot. It intentionally excludes optimizer and scheduler
+state, the private data mixture, and intermediate lineage checkpoints.
+Training regeneration uses the source pipeline, pinned external dependencies,
+and user-prepared datasets documented in the GitHub repository.
+
+Tensor-exact release verification establishes artifact identity, not complete
+paper-table reproduction. The selected 100h posttraining checkpoint uses the main
+step-3000 budget and its available PSNR matches the manuscript. Exact FDCE
+attribution remains unresolved, so the model must remain labeled
+`unverified` rather than presented as a verified reproduction of every paper
+table.
+
+The repository does not include the external base model, datasets, tokenizer,
+text encoder, SAM3, CoWTracker, rollout videos, or metric caches. Obtain each
+dependency under its own upstream access and license terms.
 
 ## Intended use and limitations
 
 These checkpoints are research artifacts for latent-action representation,
-action-conditioned video generation, and offline evaluation. They are not a
+action-conditioned video prediction, and offline evaluation. They are not a
 robot policy, planner, safety controller, or authorization for physical
-actuation. Generalization beyond the documented backbone, datasets,
-embodiment, and action convention has not been established.
+actuation. Generalization beyond the documented backbone, data, embodiment,
+and action convention has not been established.
 
 ## Licenses
 
@@ -112,9 +169,10 @@ checkpoint files are fine-tuned derivatives of an external NVIDIA base model
 and remain subject to its applicable model terms. Datasets and optional metric
 dependencies have separate licenses and access conditions. Read
 `NVIDIA_OPEN_MODEL_LICENSE.pdf`, `MODEL_LICENSE_NOTICE.md`, `NOTICE`, and
-`external_dependencies.json` before use or redistribution. The required NVIDIA
-redistribution attribution is in `NOTICE`. The `license: other` metadata is
-intentional; it does not present every artifact as Apache-2.0.
+`external_dependencies.json` before use or redistribution. The
+`license: other` metadata is intentional.
+
+Built on NVIDIA Cosmos
 
 ## Citation
 

SHA256SUMS

@@ -1,26 +1,19 @@
-0a7da4c48656b071913b9c435919962077e45af85e0c14fee8ac4c867259d7e8  CITATION.cff
+a43d7637e7707cb406cb08421075435ab03b99042ec378590d68c0ce6d077966  .gitattributes
+231a72ef8caf95da6f5b0ce7a4c80128dc02ff26c2e7edcb5d192f440519c16b  CITATION.cff
 7382b1cf711e7a7c7af9816e0e07b49b91e7149b7c6d347225f98800fcb1f1c2  LICENSE
 a618496a9171a5afa34f3861a24d1b22732ee8e516a4bdf7136a44bf8f3155b0  MODEL_LICENSE_NOTICE.md
 f124051dcb1c7a4e6d564ba623f709d42944f7112ab4d22734a37ed53f4628e0  NOTICE
 4d2fb590aa9b30c47f2058bff17291df7fa2aa0c1bd775a20703da9bb267cfab  NVIDIA_OPEN_MODEL_LICENSE.pdf
-397ee7a79bf0a79c53eda793566e95efdd4283b9f2d46429934ad5da9032dbe8  README.md
-bcbb45c90e6c518da25eea614ee2e8d76b5a985b34563d81466f4fc3f6e3dc6d  action_contract.json
-42e45fd63ee01c87e1feef21365daa79aba5864ac6c917594c45ba97fe448121  asset_manifest.json
-59643760db6ad9ed134927c07fa9c3ffaa36a10a624cdf2a968f024a681dcab0  external_dependencies.json
-96c4d31fbf7019b2f4b9ea2f19af51c8b06349328e2c153d40e668fac6b5ff8f  models/2b/1000h/stage2-step2000/model.pt
-ab27dbb3763a610b1094b414f9962c61c6bca4337c5f14ef758c9b1ba61b95d2  models/2b/1000h/stage2-step2000/model.pt.json
-b3f5a6d45453f1948f3bae8ff601117031497ab4274d2e9e0f4cc57238895f7a  models/2b/1000h/stage3-step3000/model.pt
-7e02b0f618fa39b623875839466110faeee8627bbb6887f8747a6f7992af9b11  models/2b/1000h/stage3-step3000/model.pt.json
-a23eba8e9ae66d15e8aebe3fa66108c09270c60dee1f821d1823a0b779c3242e  models/2b/100h/stage2-step2000/model.pt
-49abaf0bee0144303e3cdf2d8fe58b43d8a7a70017016e19b08c634efa6a18dc  models/2b/100h/stage2-step2000/model.pt.json
-65baa9073ea7d1cb0d355bb65364c05376ec7ccdb01e9606476b5150d68cf716  models/2b/100h/stage3-step3000/model.pt
-acc5b2c7ee43903e05156ca25277a2c4695d8d795d82ee0feb0f11cc0f0062aa  models/2b/100h/stage3-step3000/model.pt.json
-4edda56902b62bedf8e3d2a6066003eafb260d48e382525f0a05629757c4a852  models/bridge/1000h/epoch300/model.pt
-0c38e7bdef7f0cb471dee3671030b5c3c8ed5ac937e291961b96cee4b585f5d1  models/bridge/1000h/epoch300/model.pt.json
-29b74107a65f21e56ec5a8c7d21e29e501cbe79d569a3ed1454b68c5e7d64494  models/bridge/100h/epoch300/model.pt
-02653961076b1d221c8c970b359844a08c0c56786ceae9b5d0a5c91853fcaa8c  models/bridge/100h/epoch300/model.pt.json
-d91cddb3635585571677f0081ee2aad2f09de6176563a514425bfd8f07207e6c  models/lam/1000h/stage1-step180/model.pt
-8daceafa88453d11319123d889a53d5be946d761cad8197ba45e1923988c242e  models/lam/1000h/stage1-step180/model.pt.json
-2bcafc5f46416a7e633a857c141b45b5410ab356c131858b224c7c1fd33688ae  models/lam/100h/stage1-step149/model.pt
-6915c114b4d0bd73949937fcd26027413fc9632c05aa8ecc57cf0948eb31285a  models/lam/100h/stage1-step149/model.pt.json
-31da7d947a04ff52e52aff616229edf91c4498d38cc545982007740ee6a50dbf  verification_report.json
+a153f990c72c804c320e0adfea4e318fb073732a8cc5c545a23d2bd9893010bc  README.md
+c603a22042b964011fc6d4c7656c6719dd1cb23d49a559989bd1f1393ea5fe58  asset_manifest.json
+8926664198a58c688a2dc0c63ff22b72a96d74377f3e512c4f8ff08918dfb725  external_dependencies.json
+084f9b1ad1e09926c217a0fe7668e672f2e93b89314a4ff18eeed20b6b5cb6ee  models/lam/model.pt
+51eaab6889848aa7fb099afe4188ee9b72cc540ae246791090935392bc715d26  models/lam/model.pt.json
+ca9d7eb175745c591fd8259322c7c031915a021a56c5b81770aab215f0566d38  models/posttrain-100h/action_contract.json
+e3ab4ed2fe755d7e4fb127ef8f0e9d05feda859855d05f4509797f1bae301e8a  models/posttrain-100h/bridge.pt
+78d527c288638d7da0bff8f41962b27d1c3e9bbc9f5072409650e9b2444158a9  models/posttrain-100h/bridge.pt.json
+07440025c3780f907fb2b87c27263343b02e2464f3be658f07543aada088d86c  models/posttrain-100h/model.pt
+df95cfdde8424e212739d953dc8821aeb9f5f2dc5f47eecf3c0ac40c2070400e  models/posttrain-100h/model.pt.json
+8b06cebedc3fadc64b4795bd3730b29c794ba8bf38a237c5bf19e1e3aca74b9c  models/pretrain/model.pt
+b80f57650b0d94e3823fe02f72d6972687aa29cd7764e0de711a0aaf3c14669f  models/pretrain/model.pt.json
+896e4edee7ec3e8ee87334bc3ba132661ba773ca5a9e04e1a6e63cdf3b450796  verification_report.json

external_dependencies.json

@@ -6,8 +6,14 @@
       "model_license": "NVIDIA Open Model License",
       "name": "ACWM base and released baselines",
       "model_repository": "nvidia/DreamDojo",
+      "model_revision": "89d029e10816d2995d700cb8ba06f171e0504203",
+      "base_model_path": "2B_pretrain/iter_000140000/model",
+      "base_model_files": 257,
+      "base_model_bytes": 12925543625,
+      "base_model_tree_manifest_sha256": "637680cd000b3f6490cb7f256102e1f1265404d4d179cee11d53ef6e1680d526",
       "source_license": "Apache-2.0",
-      "source_repository": "https://github.com/NVIDIA/DreamDojo"
+      "source_repository": "https://github.com/NVIDIA/DreamDojo",
+      "source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff"
     },
     {
       "distribution": "gated_reference_only",

models/2b/100h/stage3-step3000/model.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:65baa9073ea7d1cb0d355bb65364c05376ec7ccdb01e9606476b5150d68cf716
-size 4303228876

models/bridge/1000h/epoch300/model.pt.json

@@ -1,122 +0,0 @@
-{
-  "action_dim": 22,
-  "base_repository": "nvidia/DreamDojo",
-  "bridge_contract_id": "agibot-alpha-22d-raw-adjacent-stride4",
-  "bytes": 639457,
-  "data_tier": "1000h",
-  "kind": "bridge",
-  "latent_dim": 32,
-  "matching_lam_sha256": "3044958c89bc06fd56154ae2172648c8eb337e3c26fcd546aeacb61d4a8e5392",
-  "model_scale": "2B",
-  "optimizer_step": 300,
-  "paper_role": "research",
-  "release_path": "model.pt",
-  "release_sha256": "4edda56902b62bedf8e3d2a6066003eafb260d48e382525f0a05629757c4a852",
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
-  "source_sha256": "bf2aded6c70c0a908c6792450941bfb91d2bf47bfe3ff28923d7a8feee229bc3",
-  "tensors": {
-    "g_state.0.bias": {
-      "dtype": "float32",
-      "shape": [
-        256
-      ]
-    },
-    "g_state.0.weight": {
-      "dtype": "float32",
-      "shape": [
-        256,
-        22
-      ]
-    },
-    "g_state.2.bias": {
-      "dtype": "float32",
-      "shape": [
-        256
-      ]
-    },
-    "g_state.2.weight": {
-      "dtype": "float32",
-      "shape": [
-        256,
-        256
-      ]
-    },
-    "g_state.4.bias": {
-      "dtype": "float32",
-      "shape": [
-        32
-      ]
-    },
-    "g_state.4.weight": {
-      "dtype": "float32",
-      "shape": [
-        32,
-        256
-      ]
-    },
-    "h_state.0.bias": {
-      "dtype": "float32",
-      "shape": [
-        256
-      ]
-    },
-    "h_state.0.weight": {
-      "dtype": "float32",
-      "shape": [
-        256,
-        32
-      ]
-    },
-    "h_state.2.bias": {
-      "dtype": "float32",
-      "shape": [
-        256
-      ]
-    },
-    "h_state.2.weight": {
-      "dtype": "float32",
-      "shape": [
-        256,
-        256
-      ]
-    },
-    "h_state.4.bias": {
-      "dtype": "float32",
-      "shape": [
-        15
-      ]
-    },
-    "h_state.4.weight": {
-      "dtype": "float32",
-      "shape": [
-        15,
-        256
-      ]
-    },
-    "statistics.action_mean": {
-      "dtype": "float32",
-      "shape": [
-        22
-      ]
-    },
-    "statistics.action_std": {
-      "dtype": "float32",
-      "shape": [
-        22
-      ]
-    },
-    "statistics.zm": {
-      "dtype": "float32",
-      "shape": [
-        32
-      ]
-    },
-    "statistics.zsd": {
-      "dtype": "float32",
-      "shape": [
-        32
-      ]
-    }
-  },
-  "verification_status": "tensor_exact"
-}

models/bridge/100h/epoch300/model.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:29b74107a65f21e56ec5a8c7d21e29e501cbe79d569a3ed1454b68c5e7d64494
-size 639457

models/lam/1000h/stage1-step180/model.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d91cddb3635585571677f0081ee2aad2f09de6176563a514425bfd8f07207e6c
-size 2839525383

models/lam/100h/stage1-step149/model.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:2bcafc5f46416a7e633a857c141b45b5410ab356c131858b224c7c1fd33688ae
-size 2839525383

models/{2b/1000h/stage2-step2000 → lam}/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96c4d31fbf7019b2f4b9ea2f19af51c8b06349328e2c153d40e668fac6b5ff8f
-size 4303228812
+oid sha256:084f9b1ad1e09926c217a0fe7668e672f2e93b89314a4ff18eeed20b6b5cb6ee
+size 2839525895

models/lam/{1000h/stage1-step180/model.pt.json → model.pt.json}

@@ -1,17 +1,24 @@
 {
+  "base_model_path": "2B_pretrain/iter_000140000/model",
+  "base_model_revision": "89d029e10816d2995d700cb8ba06f171e0504203",
   "base_repository": "nvidia/DreamDojo",
+  "base_source_repository": "https://github.com/NVIDIA/DreamDojo.git",
+  "base_source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff",
   "bridge_contract_id": null,
-  "bytes": 2839525383,
-  "data_tier": "1000h",
+  "bytes": 2839525895,
+  "checkpoint_format": "cdlam.stage1.inference",
+  "data_tier": "multi-source",
+  "historical_training_source_commit": "315712d6d01ae09533f734e67cedd51525851520",
   "kind": "stage1",
   "matching_lam_sha256": null,
   "model_scale": "2B",
-  "optimizer_step": 180,
+  "optimizer_step": 300,
   "paper_role": "research",
+  "public_runtime_repository": "https://github.com/yufanwei/CD-LAM",
+  "public_runtime_revision": "aa1e2c33367e561d1387c33fac801e9e73893af8",
   "release_path": "model.pt",
-  "release_sha256": "d91cddb3635585571677f0081ee2aad2f09de6176563a514425bfd8f07207e6c",
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
-  "source_sha256": "3044958c89bc06fd56154ae2172648c8eb337e3c26fcd546aeacb61d4a8e5392",
+  "release_sha256": "084f9b1ad1e09926c217a0fe7668e672f2e93b89314a4ff18eeed20b6b5cb6ee",
+  "source_sha256": "8f6e8e259fc701553ffa2952ef2e84b727b6f5f0176f2a72833297e0127c311f",
   "tensors": {
     "centerer.center": {
       "dtype": "float32",
@@ -5482,5 +5489,6 @@
       ]
     }
   },
-  "verification_status": "tensor_exact"
+  "verification_status": "tensor_exact",
+  "weights_only_load": "pass"
 }

action_contract.json → models/posttrain-100h/action_contract.json

@@ -12,9 +12,10 @@
   "stats_sha256": "ea3999b4141236cf41c702b6c618bc6ab9ee696a80f154eaa91efbbf88fbd72b",
   "modality_relative_path": "shared_meta/AgiBot_modality.json",
   "modality_sha256": "29fd7cec495ddeb0baa8e6d13c3d73a259a1bb2447b7afb5fe2390f1e223a4ed",
+  "metadata_source_repository": "https://github.com/NVIDIA/DreamDojo.git",
+  "metadata_source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff",
   "legacy_bridge_sha256_by_scale_family": {
-    "100h": "346710032ca50ba18876d2727981b18db60cb9e7bea1d6efdb10056bb7232fdd",
-    "1000h": "bf2aded6c70c0a908c6792450941bfb91d2bf47bfe3ff28923d7a8feee229bc3"
+    "100h": "346710032ca50ba18876d2727981b18db60cb9e7bea1d6efdb10056bb7232fdd"
   },
   "layout": [
     {"name": "left_arm_joint_position", "start": 0, "end": 7},

models/{bridge/1000h/epoch300/model.pt → posttrain-100h/bridge.pt}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4edda56902b62bedf8e3d2a6066003eafb260d48e382525f0a05629757c4a852
-size 639457
+oid sha256:e3ab4ed2fe755d7e4fb127ef8f0e9d05feda859855d05f4509797f1bae301e8a
+size 639927

models/{bridge/100h/epoch300/model.pt.json → posttrain-100h/bridge.pt.json}

@@ -1,18 +1,25 @@
 {
   "action_dim": 22,
+  "base_model_path": "2B_pretrain/iter_000140000/model",
+  "base_model_revision": "89d029e10816d2995d700cb8ba06f171e0504203",
   "base_repository": "nvidia/DreamDojo",
+  "base_source_repository": "https://github.com/NVIDIA/DreamDojo.git",
+  "base_source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff",
   "bridge_contract_id": "agibot-alpha-22d-raw-adjacent-stride4",
-  "bytes": 639457,
+  "bytes": 639927,
+  "checkpoint_format": "cdlam.action_to_latent_bridge",
   "data_tier": "100h",
+  "historical_training_source_commit": "315712d6d01ae09533f734e67cedd51525851520",
   "kind": "bridge",
   "latent_dim": 32,
   "matching_lam_sha256": "08fa663b1b9c3bb23edddfc7cd41344684f786f611fabf7d2b6618d282a4bbb9",
   "model_scale": "2B",
   "optimizer_step": 300,
   "paper_role": "research",
-  "release_path": "model.pt",
-  "release_sha256": "29b74107a65f21e56ec5a8c7d21e29e501cbe79d569a3ed1454b68c5e7d64494",
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
+  "public_runtime_repository": "https://github.com/yufanwei/CD-LAM",
+  "public_runtime_revision": "aa1e2c33367e561d1387c33fac801e9e73893af8",
+  "release_path": "bridge.pt",
+  "release_sha256": "e3ab4ed2fe755d7e4fb127ef8f0e9d05feda859855d05f4509797f1bae301e8a",
   "source_sha256": "346710032ca50ba18876d2727981b18db60cb9e7bea1d6efdb10056bb7232fdd",
   "tensors": {
     "g_state.0.bias": {
@@ -118,5 +125,6 @@
       ]
     }
   },
-  "verification_status": "tensor_exact"
+  "verification_status": "tensor_exact",
+  "weights_only_load": "pass"
 }

models/{2b/100h/stage2-step2000 → posttrain-100h}/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a23eba8e9ae66d15e8aebe3fa66108c09270c60dee1f821d1823a0b779c3242e
-size 4303228812
+oid sha256:07440025c3780f907fb2b87c27263343b02e2464f3be658f07543aada088d86c
+size 4303229324

models/{2b/100h/stage3-step3000 → posttrain-100h}/model.pt.json

@@ -1,16 +1,23 @@
 {
+  "base_model_path": "2B_pretrain/iter_000140000/model",
+  "base_model_revision": "89d029e10816d2995d700cb8ba06f171e0504203",
   "base_repository": "nvidia/DreamDojo",
+  "base_source_repository": "https://github.com/NVIDIA/DreamDojo.git",
+  "base_source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff",
   "bridge_contract_id": "agibot-alpha-22d-raw-adjacent-stride4",
-  "bytes": 4303228876,
+  "bytes": 4303229324,
+  "checkpoint_format": "cdlam.stage3.overlay",
   "data_tier": "100h",
+  "historical_training_source_commit": "315712d6d01ae09533f734e67cedd51525851520",
   "kind": "stage3",
   "matching_lam_sha256": "08fa663b1b9c3bb23edddfc7cd41344684f786f611fabf7d2b6618d282a4bbb9",
   "model_scale": "2B",
   "optimizer_step": 3000,
   "paper_role": "unverified",
+  "public_runtime_repository": "https://github.com/yufanwei/CD-LAM",
+  "public_runtime_revision": "aa1e2c33367e561d1387c33fac801e9e73893af8",
   "release_path": "model.pt",
-  "release_sha256": "65baa9073ea7d1cb0d355bb65364c05376ec7ccdb01e9606476b5150d68cf716",
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
+  "release_sha256": "07440025c3780f907fb2b87c27263343b02e2464f3be658f07543aada088d86c",
   "source_sha256": "dc96a102dae1d884fb012b705e3bbe4f840e130a24764d877b4f9e822355d0cc",
   "tensors": {
     "trainable_state.action_embedder_B_3D.fc1.bias": {
@@ -3935,5 +3942,6 @@
       ]
     }
   },
-  "verification_status": "tensor_exact"
+  "verification_status": "tensor_exact",
+  "weights_only_load": "pass"
 }

models/{2b/1000h/stage3-step3000 → pretrain}/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b3f5a6d45453f1948f3bae8ff601117031497ab4274d2e9e0f4cc57238895f7a
-size 4303228876
+oid sha256:8b06cebedc3fadc64b4795bd3730b29c794ba8bf38a237c5bf19e1e3aca74b9c
+size 4303229324

models/{2b/100h/stage2-step2000 → pretrain}/model.pt.json

@@ -1,17 +1,24 @@
 {
+  "base_model_path": "2B_pretrain/iter_000140000/model",
+  "base_model_revision": "89d029e10816d2995d700cb8ba06f171e0504203",
   "base_repository": "nvidia/DreamDojo",
+  "base_source_repository": "https://github.com/NVIDIA/DreamDojo.git",
+  "base_source_revision": "02f119b759d5c7f84a399fdeea3c6e82e7ed6cff",
   "bridge_contract_id": null,
-  "bytes": 4303228812,
-  "data_tier": "100h",
+  "bytes": 4303229324,
+  "checkpoint_format": "cdlam.stage2.overlay",
+  "data_tier": "multi-source",
+  "historical_training_source_commit": "315712d6d01ae09533f734e67cedd51525851520",
   "kind": "stage2",
-  "matching_lam_sha256": "08fa663b1b9c3bb23edddfc7cd41344684f786f611fabf7d2b6618d282a4bbb9",
+  "matching_lam_sha256": "8f6e8e259fc701553ffa2952ef2e84b727b6f5f0176f2a72833297e0127c311f",
   "model_scale": "2B",
-  "optimizer_step": 2000,
-  "paper_role": "unverified",
+  "optimizer_step": 4000,
+  "paper_role": "research",
+  "public_runtime_repository": "https://github.com/yufanwei/CD-LAM",
+  "public_runtime_revision": "aa1e2c33367e561d1387c33fac801e9e73893af8",
   "release_path": "model.pt",
-  "release_sha256": "a23eba8e9ae66d15e8aebe3fa66108c09270c60dee1f821d1823a0b779c3242e",
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
-  "source_sha256": "6f0d0fc3e356485df82370515eff523667258dc354a5ec8baa32a60a6d91a4fc",
+  "release_sha256": "8b06cebedc3fadc64b4795bd3730b29c794ba8bf38a237c5bf19e1e3aca74b9c",
+  "source_sha256": "cffba094336590c92ea3c24deb3771458ae60d2c439e191f32717d8ccd4f0826",
   "tensors": {
     "trainable_state.action_embedder_B_3D.fc1.bias": {
       "dtype": "bfloat16",
@@ -3935,5 +3942,6 @@
       ]
     }
   },
-  "verification_status": "tensor_exact"
+  "verification_status": "tensor_exact",
+  "weights_only_load": "pass"
 }

verification_report.json

@@ -1,100 +1,18 @@
 {
-  "asset_count": 8,
-  "assets": [
-    {
-      "bytes": 2839525383,
-      "id": "lam-100h-stage1-step149",
-      "path": "models/lam/100h/stage1-step149/model.pt",
-      "release_sha256": "2bcafc5f46416a7e633a857c141b45b5410ab356c131858b224c7c1fd33688ae",
-      "status": "passed"
-    },
-    {
-      "bytes": 4303228812,
-      "id": "2b-100h-stage2-step2000",
-      "path": "models/2b/100h/stage2-step2000/model.pt",
-      "release_sha256": "a23eba8e9ae66d15e8aebe3fa66108c09270c60dee1f821d1823a0b779c3242e",
-      "status": "passed"
-    },
-    {
-      "bytes": 639457,
-      "id": "bridge-100h",
-      "path": "models/bridge/100h/epoch300/model.pt",
-      "release_sha256": "29b74107a65f21e56ec5a8c7d21e29e501cbe79d569a3ed1454b68c5e7d64494",
-      "status": "passed"
-    },
-    {
-      "bytes": 4303228876,
-      "id": "2b-100h-stage3-step3000",
-      "path": "models/2b/100h/stage3-step3000/model.pt",
-      "release_sha256": "65baa9073ea7d1cb0d355bb65364c05376ec7ccdb01e9606476b5150d68cf716",
-      "status": "passed"
-    },
-    {
-      "bytes": 2839525383,
-      "id": "lam-1000h-stage1-step180",
-      "path": "models/lam/1000h/stage1-step180/model.pt",
-      "release_sha256": "d91cddb3635585571677f0081ee2aad2f09de6176563a514425bfd8f07207e6c",
-      "status": "passed"
-    },
-    {
-      "bytes": 4303228812,
-      "id": "2b-1000h-stage2-step2000",
-      "path": "models/2b/1000h/stage2-step2000/model.pt",
-      "release_sha256": "96c4d31fbf7019b2f4b9ea2f19af51c8b06349328e2c153d40e668fac6b5ff8f",
-      "status": "passed"
-    },
-    {
-      "bytes": 639457,
-      "id": "bridge-1000h",
-      "path": "models/bridge/1000h/epoch300/model.pt",
-      "release_sha256": "4edda56902b62bedf8e3d2a6066003eafb260d48e382525f0a05629757c4a852",
-      "status": "passed"
-    },
-    {
-      "bytes": 4303228876,
-      "id": "2b-1000h-stage3-step3000",
-      "path": "models/2b/1000h/stage3-step3000/model.pt",
-      "release_sha256": "b3f5a6d45453f1948f3bae8ff601117031497ab4274d2e9e0f4cc57238895f7a",
-      "status": "passed"
-    }
-  ],
-  "checks": {
-    "asset_manifest_and_release_sha256": "passed",
-    "checkpoint_top_level_schema": "passed",
-    "cjk_scan": "passed",
-    "legacy_experiment_identifier_scan": "passed",
-    "nvidia_agreement_copy_identity": "passed",
-    "nvidia_cosmos_attribution_presence": "passed",
-    "nvidia_redistribution_notice_presence": "passed",
-    "private_machine_path_scan": "passed",
-    "sanitizer_reload_tensor_comparison": "tensor_exact",
-    "symlink_scan": "passed",
-    "upstream_branding": "allowlisted_attribution_only",
-    "weights_only_checkpoint_load": "passed"
-  },
-  "nvidia_agreement": {
-    "authoritative_source": "https://www.nvidia.com/content/dam/en-zz/Solutions/license-agreements/enterprise-software/nvidia-open-model-license-agreements-24-10-2025.pdf",
-    "path": "NVIDIA_OPEN_MODEL_LICENSE.pdf",
-    "sha256": "4d2fb590aa9b30c47f2058bff17291df7fa2aa0c1bd775a20703da9bb267cfab",
-    "version_release_date": "October 24, 2025"
-  },
-  "release_family": "CD-LAM-2B",
+  "auxiliary_bytes": 641320,
+  "auxiliary_file_count": 2,
+  "checkpoint_inventory_closed": true,
+  "checkpoint_schema_validation": "pass",
+  "distribution_scope": "inference_evaluation",
+  "hash_validation": "pass",
+  "license_inventory": "pass",
+  "main_model_bytes": 11445984543,
+  "main_model_count": 3,
+  "manifest_sha256": "c603a22042b964011fc6d4c7656c6719dd1cb23d49a559989bd1f1393ea5fe58",
+  "private_marker_scan": "pass",
+  "release_id": "cd-lam-2b-three-entry",
   "schema_version": 1,
-  "source_code_commit": "315712d6d01ae09533f734e67cedd51525851520",
-  "text_files_with_allowlisted_upstream_attribution": [
-    "MODEL_LICENSE_NOTICE.md",
-    "asset_manifest.json",
-    "external_dependencies.json",
-    "models/2b/1000h/stage2-step2000/model.pt.json",
-    "models/2b/1000h/stage3-step3000/model.pt.json",
-    "models/2b/100h/stage2-step2000/model.pt.json",
-    "models/2b/100h/stage3-step3000/model.pt.json",
-    "models/bridge/1000h/epoch300/model.pt.json",
-    "models/bridge/100h/epoch300/model.pt.json",
-    "models/lam/1000h/stage1-step180/model.pt.json",
-    "models/lam/100h/stage1-step149/model.pt.json"
-  ],
-  "upstream_base_references_in_checkpoint_metadata": 8,
-  "verification_status": "passed",
-  "weight_bytes": 22893245056
+  "status": "upload_ready",
+  "tensor_exact_staging": "pass",
+  "weights_only_load": "pass"
 }