Separate LAW WAM from its complete VLM package

README.md

@@ -2,62 +2,15 @@
 license: apache-2.0
 tags:
 - robotics
-- law
+- LAW
+- world-action-model
 ---
+# LAW-OpenWAM
 
-# LAW-OpenWAM-20D
+World-action model for **LAW + OpenWAM**, selected step-120,815 checkpoint. Use with the complete [LAW-VLM-OpenWAM](https://huggingface.co/KAITANG2003/LAW-VLM-OpenWAM) Q1/Q2/Q3 VLM.
 
-Final 20D policy: OpenWAM step 120815 and Q3 step 50000. Uses the shared Q1/Q2 repository and the EEF20/unified80 conversion. Historical full evaluation: 95.56%; this historical benchmark was not rerun.
+Includes WAM weights, configuration and normalization statistics under `wam/`. The corresponding VLM is packaged separately. Original combined releases remain available in this repository's Git history.
 
-## Verified training and inference
+Code and download guide: [hey-cjj/LAW](https://github.com/hey-cjj/LAW). Historical RoboTwin overall success: 95.56%. These are selected historical results, not a new benchmark. Six bounded 32-GPU training save/load checks and both bounded real RoboTwin policy episodes passed before this packaging change; see [release audit](https://github.com/hey-cjj/LAW/blob/main/docs/release_audit.md) for scope and historical provenance limits.
 
-Validation completed on 2026-10-12 with the matching scoped LAW code.
-The tested code revision is `26141a69e32ad40b89eb765d4eeb6990ce013c2c`;
-subsequent documentation commits may retain the same tested numerical code.
-Matching code: [hey-cjj/LAW](https://github.com/hey-cjj/LAW). The tested code snapshot is retained in its Git history; see the repository checkpoint-binding and release-audit documents.
-
-All six selected training entry points ran on four nodes with eight A800 GPUs
-each (32 GPUs): FastWAM16D, OpenWAM20D, Q1, Q2, Q3-16D and Q3-20D.
-Each performed at least two actual optimizer updates and saved a nonempty,
-loadable checkpoint. Recorded losses and saved model tensors were finite.
-Q2 initialized from the Q1 checkpoint produced in that test. Both Q3 tests
-initialized from the selected historical shared Q1/Q2 model. The original
-batch/lr parameters, memory mode and ZeRO-1 settings were retained; only test
-length, checkpoint frequency and FastWAM loss logging frequency were shortened.
-The OpenWAM Accelerate 1.0.1 gradient-norm logging fix was tested in real training.
-VLM tests used fresh container-local environments. WAM fresh package installs
-were checked; the final repaired WAM tests reused verified isolated shared
-role environments without changing their packages.
-
-The selected final checkpoints passed matching loader/key/shape and bounded
-real inference checks for WAM, Q1/Q2 and both Q3 critics. Both complete production
-policy pipelines then completed one RoboTwin episode successfully on
-`pick_dual_bottles/demo_clean`, episode 0, fixed seed `4300000` (FastWAM 1/1,
-OpenWAM 1/1), including three-camera RGB/depth rendering, cuRobo planning,
-Q1 generation, Q3 scoring and action execution. Both used 32 candidates,
-8 elites, micro-batch 32, and fixed Q2 done1 behavior with threshold 0.5.
-That episode produced one Q1 subtask; Q2 scores were checked separately and
-this episode is not evidence of a Q2-triggered subtask transition.
-
-These are real launch/update/save and inference/evaluation validations.
-They do not establish full retraining, a new full benchmark, or byte-for-byte
-historical reproduction. Existing latent caches were used for short training;
-no latent cache was rebuilt or uploaded. The cache-generation code is provided,
-but raw-download-to-new-cache-to-training was not rerun end to end.
-
-## Checkpoint binding and download
-
-Five final groups are packaged across `LAW-16D` (WAM/Q3-16D), `LAW-OpenWAM-20D` (WAM/Q3-20D) and `LAW-Q1Q2` (shared VLM). Both WAMs use `first_layer_only` + `subtask` memory. Initialize training using the separate ActionDiT/Qwen/OpenWAM Alpha dependencies.
-
-Use the release downloader, `python scripts/download.py models --profile all`, to verify hashes and relocate original metadata. The original configuration snapshots may contain historical paths; the release prepares local configuration paths without altering tensors. Models remain private: authenticate using HF login or HF_TOKEN.
-
-```python
-from huggingface_hub import snapshot_download
-snapshot_download("KAITANG2003/LAW-OpenWAM-20D", local_dir="checkpoints/LAW-OpenWAM-20D")
-```
-
-See `MODEL_FILES.json` for original file sizes and SHA256. Historical Git/dirty-state and optimizer/RNG limits are recorded in the matching code release's `docs/checkpoint_binding.md`.
-
-## License
-
-LAW-authored weights are Apache-2.0. Upstream licenses/notices are retained. See [LICENSE_WEIGHTS.txt](LICENSE_WEIGHTS.txt) and [LICENSE_SCOPE.md](LICENSE_SCOPE.md) for the full terms and scope. Code, annotations and original media have separate licensing.
+Tensor bytes are unchanged. LAW-authored trained weights/adapters are Apache-2.0; upstream materials retain their terms. See `LICENSE_WEIGHTS.txt` and `LICENSE_SCOPE.md`. Models remain private.

q3/q3_config.json

@@ -1,43 +0,0 @@
-{
-  "model": {
-    "base_model": "/mnt/liying_ckp/qwen_ft/Qwen3-VL-2B-Instruct",
-    "source_checkpoint": "/mnt/liying_ckp/qwen_ft/robotwin_qwen3vl_2b-fastwam-stage2-q2-lora_20260830_140608/checkpoint-50000",
-    "q3_init": "q2_copy",
-    "q3_lora_rank": 128,
-    "q3_lora_alpha": 256,
-    "q3_lora_target_modules": [
-      "q_proj",
-      "k_proj",
-      "v_proj",
-      "o_proj"
-    ]
-  },
-  "data": {
-    "sft_data_root": "/mnt/dataset/pku-data/robotwin_gt_depth_ee",
-    "failure_data_root": "/mnt/dataset/pku-data/robotwin_failure_semantic_final_v2_train_v2",
-    "exclude_sft_episodes": "/mnt/world_foundational_model/kai/dev/CogWAM/qwen-vl-progress-estimation/qwen-vl-finetune/config/robotwin_test_episodes.json",
-    "failure_fraction": 0.25,
-    "action_dt": 0.06,
-    "horizon": 32,
-    "action_dim": 20
-  },
-  "training": {
-    "output_dir": "/mnt/kai_ckp/q3/robotwin_qwen3vl_2b-fastwam-stage3-q3-subtask-q2copy-attnonly-r128-20d-20260930/checkpoints",
-    "max_steps": 50000,
-    "per_device_train_batch_size": 8,
-    "gradient_accumulation_steps": 1,
-    "dataloader_num_workers": 4,
-    "learning_rate": 1e-05,
-    "weight_decay": 0.0,
-    "warmup_steps": 1000,
-    "save_steps": 2000,
-    "model_max_length": 8192,
-    "bf16": true,
-    "gradient_checkpointing": true,
-    "gradient_clip_norm": 1.0,
-    "seed": 42,
-    "num_nodes": 4,
-    "gpus_per_node": 8,
-    "expected_world_size": 32
-  }
-}

q3/q3_state.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:edae85178479b37360c3e62fe4f4ffff960086e9afcbfd9f55e1160f7c5bff10
-size 241345455

q3/source_checkpoint.json

@@ -1,16 +0,0 @@
-{
-  "source_checkpoint": "/mnt/liying_ckp/qwen_ft/robotwin_qwen3vl_2b-fastwam-stage2-q2-lora_20260830_140608/checkpoint-50000",
-  "q3_init": "q2_copy",
-  "q3_lora_rank": 128,
-  "q3_lora_alpha": 256,
-  "q3_lora_target_modules": [
-    "q_proj",
-    "k_proj",
-    "v_proj",
-    "o_proj"
-  ],
-  "initialized_from_q3": null,
-  "reward_mode": "subtask",
-  "gt_reward_mode": "subtask",
-  "rollout_reward_mode": "subtask"
-}