| --- |
| license: other |
| pretty_name: BEHAVIOR Explore Legacy Artifacts |
| --- |
| |
| # BEHAVIOR Explore legacy archive |
|
|
| This public dataset repository preserves audit evidence from retired RPent / RLinf BEHAVIOR prompt-following routes. Every artifact under `legacy/` is historical evidence only. |
|
|
| No legacy artifact—including `pi05-b1kpt50-cs32`, canonical steps 365/767/1167, or native-LoRA step338—may initialize, resume, warm-start, merge into, or select a checkpoint for the current LA/VLA route. Successful archival or restoration does not change this policy. The current route starts separately from the official Physical Intelligence `pi05_base` checkpoint. |
|
|
| ## Published content and status |
|
|
| | Path | Publication status | Integrity status | Meaning | |
| |---|---|---|---| |
| | `legacy/models/pi05-b1kpt50-cs32/` | uploaded | covered by the original archive record | Retired third-party BEHAVIOR pt50 model. | |
| | `legacy/training/canonical/` | uploaded | not exhaustively enumerated by the LoRA manifest | Retired canonical checkpoints, deploy exports, validation receipts, and logs. | |
| | `legacy/training/lora/` | uploaded | 21/21 source files verified at the fixed payload revision; see `manifests/archive-manifest-v2.json` | Retired native-LoRA run, sealed step338 DCP, associated CPU-validation evidence, TensorBoard event, and driver log. | |
| | `legacy/training/dataset-deltas/obsolete-prompt-following-derived-v1/` | uploaded | 7/7 delta files plus reconstruction manifest verified at fixed payload revision | Byte deltas and reconstruction recipes for five deleted, obsolete derived trees; not current training data. | |
| | `legacy/diagnostics/` | not uploaded | not verified | Planned destination for interrupted campaigns; directory presence must not be inferred from this README. | |
| | `legacy/source-snapshot/` | not uploaded | not verified | Planned destination for selected source/config snapshots; directory presence must not be inferred from this README. | |
|
|
| Only files explicitly enumerated in an active manifest are covered by that manifest's integrity claims. A path named in this README or a manifest is not, by itself, evidence that the path was uploaded. |
|
|
| ## Native-LoRA step338 |
|
|
| The complete retained local LoRA source tree was uploaded without rewriting its bytes to `legacy/training/lora/` in payload revision `d6ca4dad7cf9a2b897a09c7dd29a41b3aff62141`: |
|
|
| - 21 regular files; |
| - 16,978,218,437 bytes; |
| - sealed DCP object `step-000000000338-1f2c404a246ebabc`; |
| - DCP object manifest SHA-256 `1f2c404a246ebabc77215999c898a7a848cf3f0656834e4962adcae08864c5dd`; |
| - associated 32-forward CPU validation receipt with `masked_val_flow_loss=0.007970079779624939`. |
|
|
| The associated CPU receipt originally validated a temporary BF16 model export. It must not be described as proof that the archived DCP training state was restored. The manifest separately records source integrity, remote per-file verification, isolated re-download verification, model-state DCP decoding, and production-equivalent resume status. |
|
|
| The archived `.pt`, `.distcp`, and DCP metadata files are trusted checkpoint inputs that may use pickle-based serialization. Load them only in a trusted, isolated environment after verifying the manifest. Original logs and receipts intentionally preserve machine-local provenance such as absolute paths, hostnames, PIDs, and timestamps. A bounded scan of the text, config, JSON, and log files found no token, API-key, private-key, password, cookie, or authorization-header pattern; this is not a general claim about arbitrary binary checkpoint contents. |
|
|
| After the fixed-revision remote verification, isolated re-download, and CPU DCP model-state decode had passed, the old local LoRA copies were deleted under an explicit user instruction. The deletion covered only two verified roots: `/home/ubuntu/lwb/Artifacts/RLinf_prompt_following/lora` (21 regular files; 16,978,218,437 regular-file bytes) and its byte-matched checkpoint/validation copy `/dev/shm/rpent_prompt_following_lora_native_lora_v1` (18 regular files; 16,977,337,528 regular-file bytes). The first root was deleted at `2026-08-13T13:16:52+08:00` and the second at `2026-08-13T13:22:02+08:00`. Their deletion receipt is `manifests/verification/native-lora-step338-local-deletion.json`. Production-equivalent eight-rank resume remains `not_run`; the user-authorized deletion does not turn that missing evidence into a pass. |
|
|
| ## Obsolete derived dataset deltas |
|
|
| Payload revision `5c40e01b9fec4d54f15af1a98bff367ce77a6a35` stores seven unique delta files totaling exactly 61,711 bytes plus `reconstruction-manifest-v1.json` as additional metadata. The deltas cover five obsolete derived roots: canonical variants A and B, converter outputs run1 and run2, and the non-canonical task-0001 v1 tree. |
|
|
| At that fixed revision, all 8/8 payload paths, sizes, and SHA-256 values were verified after an isolated download. The five complete roots were then reconstructed in isolation and matched their still-present sources: 43 regular files and 14,152,687,590 bytes, with exact path sets and per-root tree-contract SHA-256 values. The five source roots were deleted only after those checks passed; see `manifests/verification/obsolete-prompt-following-derived-v1-local-deletion.json`. |
|
|
| This is a delta archive, not a self-contained dataset release and not an input for the current LA/VLA route. Reconstruction depends on the retained, hash-pinned canonical-v1-c and task-0001 canonical bases declared in the reconstruction manifest. Those bases are not duplicated by this delta archive and remain outside the deletion scope. |
|
|
| ## Data and campaign boundaries |
|
|
| - The canonical task-0001 dataset contains 142 train episodes, 1,704 segments, and 663,708 canonical 32-step windows. It remains local training input and is not duplicated here. |
| - The stopped `rpent-anchor-272-v2` snapshot recorded 70 cases with completed status, 16 stale-running status files, and 186 unstarted cases. “Completed status” is not task success. No final aggregate exists, so no formal completion or success rate may be derived from it. |
| - `exact_historical_restore=false` for Parquet anchor materialization receipts; one explicit warm-up physics tick was part of the accepted state-construction contract. |
|
|
| ## Verification vocabulary |
|
|
| These states are independent: |
|
|
| - `uploaded`: files are present at a fixed Hub revision. |
| - `remote_per_file`: remote path, size, and SHA-256 match the source manifest. |
| - `isolated_redownload`: a new fixed-revision download matches the source manifest byte-for-byte. |
| - `dcp_model_state_decode`: the downloaded DCP was parsed and its model state was materialized on CPU. |
| - `production_equivalent_resume`: model, optimizer, scheduler, per-rank data state, and RNG were restored in an isolated matching distributed environment and advanced through a controlled resume step. |
|
|
| Do not promote a weaker state to a stronger one. In particular, CPU loss, local predicates, materialization success, or checkpoint presence is not robot task success. |
|
|
| ## Guidance for future agents |
|
|
| 1. Use `manifests/archive-manifest-v2.json` for the LoRA file list and verification state; retain v1 only as historical evidence. |
| 2. Do not delete a local source merely because files are visible on the Hub. Require an exact intended deletion scope, fixed-revision remote verification, isolated re-download, explicit deletion authorization, and a receipt that records both completed checks and any remaining restore gap. |
| 3. Do not mix interrupted v1/v2 campaign receipts or describe partial campaigns as complete datasets. |
| 4. Keep official task success, local predicates, workflow sealing, terminal receipts, and publication as separate facts. |
|
|
