| --- |
| library_name: transformers |
| tags: |
| - robotics |
| - reward-modeling |
| - qwen3-vl |
| - peft |
| - offline-training |
| --- |
| |
| # Cosmos–Robometer Phase 2.1A training code |
|
|
| Portable training kit for the approved `Public-Proxy stage2_core_v1` |
| experiments: |
|
|
| - `Robometer-Low-LoRA`: implemented formal trainer with frozen ID-dev evaluation, |
| early stopping, and a 3,000-step hard cap. |
| - `Robometer-Low`: full language-decoder smoke test. Formal training remains |
| intentionally blocked until suitable multi-GPU/server memory is available. |
|
|
| The repository does not duplicate the official Robometer trajectory archives, |
| model weights, checkpoints, credentials, OOD evaluation archives, or |
| machine-specific run logs. It does contain derived action sidecars and their |
| manifests: 20,620 sidecar files are packaged in a 455,188,480-byte tar archive. |
| These sidecars are not consumed by the Phase 2.1A reward-model training loop. |
|
|
| ## Recreate the environment |
|
|
| ```bash |
| uv python install 3.10.12 |
| uv sync --project environment/phase21a_robometer --frozen |
| uv run --project environment/phase21a_robometer \ |
| python environment/phase21a_robometer/doctor.py |
| ``` |
|
|
| The environment is locked for PyTorch 2.8.0/CUDA 12.8. The reference smoke |
| hardware was one RTX 5090 with 32,607 MiB VRAM. |
|
|
| For a server with restricted or no network access, download the validated pure |
| uv offline archive under `offline_env/`. It includes uv 0.11.3, CPython 3.10.12, |
| all 50 locked packages, CUDA 12.8 Python wheels, and the locally built |
| FlashAttention wheel. See [OFFLINE_UV_DEPLOYMENT.md](OFFLINE_UV_DEPLOYMENT.md). |
|
|
| ## External inputs |
|
|
| - Base model: `unsloth/Qwen3-VL-4B-Instruct` at revision |
| `252d592b59b0233b226875a44ac135cfa1d3f755`. |
| - Dataset route: `Public-Proxy stage2_core_v1` derived from |
| `robometer/processed_datasets` at revision |
| `1a8eb2d279c70964fadeaf8019a77c490c3abe02`. |
| - Frozen Phase 2.1A subset: 2,689 pair rows referencing 4,323 trajectory files. |
|
|
| There is no separate `ar-mine/cosmos-robometer-public-dataset` repository. The |
| subset is reconstructed from the pinned official Robometer processed-data |
| repository and the frozen dataset manifest in this repository. |
|
|
| ## Training |
|
|
| See [TRAINING_MANUAL.md](TRAINING_MANUAL.md). The frozen LoRA configuration is |
| `configs/phase21a_robometer_low_lora_train_v2.yaml`. |
|
|
| For the pinned official `stage2_core_v1` download, extraction, frozen-manifest |
| association, and optional recovered-action linkage, see |
| [DATA_SETUP.md](DATA_SETUP.md). |
|
|
| ## Current support boundary |
|
|
| The LoRA trainer is runnable and validated at micro-batch 32, gradient |
| accumulation 2, effective batch 64. The full-tune file is a capacity smoke test, |
| not a formal distributed trainer. Paged 8-bit AdamW is not permitted for the |
| formal `Robometer-Low` comparison. |
|
|
| ## Redistribution |
|
|
| This public repository includes experiment code, metadata, and recovered action |
| sidecars, but does not grant redistribution rights for upstream datasets or |
| model weights. Review the terms of each upstream source before reusing or |
| redistributing derived assets. |
|
|
