Skip to content

Model card template

Copy everything below the line into the checkpoint repo's README.md and replace every <placeholder>. Remove the ``yaml fence lines so the metadata block starts with---` on line 1 (the Hub only parses unfenced frontmatter). Keep the section order. Follow the writing-style rules in model-cards.md: no em dashes in prose, no rhetorical bold, no decorative emoji. Rules and the gating checklist: model-cards.md.

---
library_name: <lerobot | transformers | ...>
pipeline_tag: robotics
license: <inherited license, or `other` + license_name + license_link>
base_model: <org/base-model>
datasets:
  - <org/dataset>
tags:
  - robotics
  - vla
  - <robot>
  - <method>
inference: false
# RECOMMENDED:
# co2_eq_emissions: {emissions: <g>, source: "<estimator>", training_type: "fine-tuning",
#                    geographical_location: "<region|unknown (...)>", hardware_used: "<n x GPU>"}
# model-index: [{name: <repo>, results: [{task: {type: robotics},
#   dataset: {type: <id>, name: <name>},
#   metrics: [{type: <metric>, name: <display>, value: <n>}]}]}]
---

Evaluate with Inspect Robots Adapters: <embodiment> Benchmarks: WorldEvals Catalog: WorldPolicies

<≤3 sentences: base model + method (LoRA/full FT; what's frozen) + robot + dataset. Adapter/merge status and where raw adapters live.>

Intended use & safety

  • Intended use: research/evaluation on for .
  • Out of scope: other embodiments/rigs without fine-tuning (no zero-shot embodiment transfer); unattended operation; operation near people without hardware e-stop and enforced workspace/torque limits.
  • Validation status: (). Users are responsible for safe integration (guardrails, e-stop, workspace limits) before any deployment.

Training
Data
Embodiment
Image preprocessing
Schedule
Checkpoint selection
Headline curve : →

Losses & evaluation

  • Training loss: ; trainable: <...>; frozen: <...>.
  • Eval regime: .
  • Comparisons: .
  • Scope: .

Provenance
Trained by
Training code (private): \">
Framework Z>
Compute provider
Wall-clock <hours; preemptions: , max steps lost
Total compute <FLOPs or GPU-hours + method + uncertainty
Cost <USD run / USD project-total
Experiment tracking <platform, project, run ids
Card authorship " vs first-hand>

Usage

<zero-edit loading snippet>
  • Observations: cameras ; state ().
  • Actions: , , horizon steps<, control ~ Hz if known>. Normalization statistics: .
  • Serving: .

Data provenance & caveats

Versioning & contact

  • <main stability statement; pin by revision for reproduction<; new_version chain note>.>
  • Issues: .