bageldotcom
/

Bagel-Robotics-1

@@ -8,7 +8,7 @@ tags:
 - robodojo
 ---
-# VPP2 compact recovery — update 1,000
 Complete: listed model artifacts verified against local receipts and remote checksums.
@@ -37,18 +37,22 @@ The original `COMPLETE.json` has `simulation_completed: false`, reflecting its e
 Authenticate with an account that has access to this private repository, then run:
 ```bash
-hf download GnanaPraveen/vpp2-recovery-step-1000 --local-dir /data3/praveen/models/vpp2-recovery-step-1000
 ```
 Use the code and setup link below to load these files. Keep the integrity receipts with the downloaded weights.
 ## Lineage and code
 Mask learning update 1,500 selected the channels. The physically pruned video expert has **12,468,409,664 parameters**, compared with 16,394,878,784 before pruning (23.95% fewer video parameters). The frozen Action2B expert is separate from this count. Recovery checkpoint 55 already had 55 recovery updates with K/V supervision; the later no-K/V branch continued from it to update 1,000.
-[Exact code and setup](https://github.com/bageldotcom/bagel-robotics/tree/2fc74c246136e543d72ff9af1732936814d3baf8/workflows/distillation) · [Integration PR](https://github.com/bageldotcom/bagel-robotics/pull/169)
-These artifacts require the repository's compact-aware VPP2 loader; they are not Transformers `AutoModel` checkpoints. Follow `docs/MASK_COMPRESSION.md` and the refined package setup instructions in that code snapshot. External VAE/conditioning encoders and RoboDojo data remain separate dependencies. Original upstream model/data terms continue to apply; this backup does not assert a new redistribution license.
 ## Integrity and training continuation

 - robodojo
 ---
+# AdaPrune — Bagel-Robotics-1 checkpoint 1000
 Complete: listed model artifacts verified against local receipts and remote checksums.
 Authenticate with an account that has access to this private repository, then run:
 ```bash
+hf download bageldotcom/adaprune \
+  --revision 2651631c5d08819088475e8778676f5cf6d9afcf \
+  --local-dir /absolute/path/to/checkpoints/cp1000
 ```
 Use the code and setup link below to load these files. Keep the integrity receipts with the downloaded weights.
+For the prepared XPolicyLab submission, the policy directory is `policy/bagel_robotics_1` and the checkpoint name is `cp1000`. Its pinned downloader fetches the six required bundle files and the separate VAE/conditioning assets, verifies their hashes, and supplies the standalone inference runtime. The adapter PR is being prepared; no public XPolicyLab PR link is available yet. The technical report and training/pruning/export code are planned for November 20, 2026. Evaluator access must be arranged while this repository remains private.
 ## Lineage and code
 Mask learning update 1,500 selected the channels. The physically pruned video expert has **12,468,409,664 parameters**, compared with 16,394,878,784 before pruning (23.95% fewer video parameters). The frozen Action2B expert is separate from this count. Recovery checkpoint 55 already had 55 recovery updates with K/V supervision; the later no-K/V branch continued from it to update 1,000.
+[Historical research code (access may be restricted)](https://github.com/bageldotcom/bagel-robotics/tree/2fc74c246136e543d72ff9af1732936814d3baf8/workflows/distillation) · [Historical research integration PR](https://github.com/bageldotcom/bagel-robotics/pull/169)
+These artifacts require a compact-aware VPP2 loader; they are not Transformers `AutoModel` checkpoints. The prepared XPolicyLab adapter includes that inference runtime and does not require access to the historical research repository. Until its PR is available, reviewers need the standalone adapter package from Bagel Labs. External VAE/conditioning encoders and RoboDojo data remain separate dependencies. Original upstream model/data terms continue to apply; this backup does not assert a new redistribution license.
 ## Integrity and training continuation