srb model — Manage Model Artifacts
srb model packages checkpoints as SRB artifacts with identity, optional
metrics, dataset provenance, and framework restore metadata.
Usage
srb model <action> [options]
Actions
| Action | Purpose |
|---|---|
create CHECKPOINT --output OUT | Package one checkpoint into a model artifact. Requires --algo; task, embodiment, and action mode can be inferred from --dataset. |
from-run RUN --output OUT | Package the latest checkpoint from a completed SRB run. The algorithm can be inferred from metadata.json. |
info PATH | Print the model manifest as JSON. |
validate PATH | Validate the artifact manifest, checkpoint, and bundled config paths. Use --strict to treat warnings as errors. |
card PATH | Write a Hugging Face README.md model card. |
push PATH | Upload or dry-run a Hugging Face model publication. |
create and from-run accept repeated --metric key=value and
--metadata key=value. Metrics must be finite numbers; metadata values are
strings.
Portable Restore
from-run preserves framework restore context. Dreamer and TD-MPC2 can include
config.yaml. RSL-RL first uses a run-level config.yaml when present;
otherwise RSL-RL and SKRL use the shared fallback order: agent from
.hydra/config.yaml, then the run.agent_cfg snapshot in metadata.json.
When the metadata.json snapshot is the fallback, packaging fails if it holds
redacted or non-finite values. Credential-bearing keys and strings are removed with a warning that
names only key paths. Artifact directories can be passed wherever checkpoint
paths are accepted by srb agent eval, collect, teleop, and
srb real_agent eval or validate.
SB3/SBX artifacts preserve RL-Zoo normalization state. Normalized runs require
vecnormalize.pkl; inference restores its frozen statistics without
normalizing evaluation rewards. Missing or contradictory context fails
closed. Older artifacts without a normalization declaration need repackaging
for SB3/SBX inference. Dict observations and recurrent state keep the trained
layout.
Publication requires dataset provenance from --dataset, a consumed-demo run,
or --dataset-repo-id with optional --dataset-revision.
Preparation and validation stream hashes under per-file and aggregate byte limits. Validation rejects files that grow or change while being read.
Legacy v1/v2 restore verifies hashes while copying files into owned temporary storage and removes partial copies on failure. Large checkpoints can exceed the 256 MiB whole-byte API limit within the streamed file and artifact limits; framework deserialization limits still apply.
Examples
# Package the latest checkpoint from a run and infer identity from the dataset
srb model from-run ${SRB_LOGS_DIR}/excavation/rsl_rl_ppo/<run> \
--dataset /data/srb_excavation_policy \
--output /models/srb_excavation_rsl_rl_ppo
# Validate and inspect the artifact
srb model validate /models/srb_excavation_rsl_rl_ppo --strict
srb model info /models/srb_excavation_rsl_rl_ppo
# Dry-run the Hugging Face upload, including the generated README.md
srb model push /models/srb_excavation_rsl_rl_ppo --dry-run
# Reuse the packaged artifact for policy rollout
srb agent collect --env excavation --model /models/srb_excavation_rsl_rl_ppo