What is the Model Onboard skill?
Model Onboard is a ai engineering skill by NVIDIA for Claude Code, Cursor, Codex and other AI coding agents. Implement an upstream pretrained model natively in BioIR using its optimized components, with checkpoint conversion and numerical validation. Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.
Input: upstream source and revision, a pretrained checkpoint, a writable BioIR checkout, and the requested model features. Output: a native BioIR implementation, checkpoint conversion, tests and validation evidence, and data pipeline integration when in scope.
When to use it
Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.
What the SKILL.md covers
- Establish the Scope
- Execution and Evidence
- Report
- Author
- NVIDIA
- Pack
- NVIDIA/skills (30 skills)
- Category
- AI engineering, #99 of 610
- Alyn Score
- 78 / 99
- Files
- 17, includes scripts
- License
- Apache-2.0
- Last updated
- 4 h ago
- Works with
- Claude Code, Claude, Cursor, Codex, Gemini CLI, Copilot
Model Onboard compared to similar skills
| Skill | Alyn Score | GitHub stars | Updated |
|---|---|---|---|
| Model Onboard · NVIDIA | 78 | 4.1k | 4 h ago |
| Architecture Zoo · Aperivue | 44 | 388 | 3 days ago |
| Implement Spec · vinvcn | 68 | 5.9k | 2 days ago |
| Implement Spec · Matt Pocock | 88 | 328k | 26 h ago |
Install Model Onboard
Pick your agent. Every command is ready to paste.
npx alyn-cli add model-onboardDetects your agents (Claude Code, Cursor, Codex and 70+ more) and installs to the right folder. Add -g to install globally.
npx skills add NVIDIA/skills --skill model-onboardOr use the open skills CLI directly.
01
Install it
Run the command above, or upload the folder in Claude.
02
Ask naturally
Describe the task. Your agent loads model-onboard when the request matches its description.
03
Name it if needed
If it does not trigger, say “use the model-onboard skill”. In Codex, type $model-onboard.
Model Onboarding
Input: upstream source and revision, a pretrained checkpoint, a writable BioIR checkout, and the requested model features. Output: a native BioIR implementation, checkpoint conversion, tests and validation evidence, and data pipeline integration when in scope.
When the request is an offline decision about a recorded input, write that decision and stop. Do not clone, implement, or run a model.
The procedure lives in the BioIR checkout, written $BIOIR below. Read
$BIOIR/docs/advanced/port-model.md in full and follow it. Without a
checkout, clone https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime
and work on a branch. The guides and the bench-perf-oss skill named below
are in that checkout, so stop and report if you can reach neither a checkout
nor the clone. Reuse BioIR's modules where their math matches
upstream, and implement the architecture's differences. An adapter that
only swaps one module into the upstream model does not complete a native
model port; for that task, follow
$BIOIR/docs/advanced/accelerate-custom-model.md instead.
Establish the Scope
Use the source, checkpoint, checkout, and working directory already supplied in the conversation. Ask only for missing information that affects the port: the authoritative upstream revision and checkpoint, the model variant, and whether full data pipeline integration is required.
Work on a branch in the BioIR checkout. Keep upstream read-only and its dependencies in a separate environment. Store captures, reference outputs, validation scripts, and reports in an external working directory; keep reusable implementation and tests in BioIR.
Before implementing, write a feasibility report in the working directory. Map the upstream components to BioIR components, explain mathematical and layout differences, list missing operations and input types, and state the implementation and validation scope. Resolve an ambiguous architecture or checkpoint with the user; routine implementation choices can proceed.
Execution and Evidence
- Inspect the available hardware before scheduling validation. Without a supported GPU, continue source analysis, implementation, and checks that run on CPU. Save the GPU validation commands and mark those checks pending; never report the port as validated from static checks alone.
- Load checkpoints with
weights_only=True. Account for checkpoint parameters and buffers: fail on a tensor that is unread or read twice, and explain intentionally unused tensors and shared weights. Never hide missing mappings with non-strict loading. - Establish module equivalence on identical captured inputs before relying on end-to-end quality scores, then compare the whole model with the upstream's spread between repeated runs. Record upstream and BioIR revisions, checkpoint identity, settings, tolerances, commands, and artifacts.
- When full pipeline support is requested, follow
$BIOIR/docs/advanced/port-data-pipeline.mdand$BIOIR/docs/advanced/validate-data-pipeline.md. Reuse an existing pipeline only after checking its feature and output semantics against upstream. - Benchmark only after correctness checks pass. For supported folding
models, use the
bench-perf-ossskill at$BIOIR/skills/bench-perf-oss/SKILL.md; retain its scope and applicability rules.
Report
State what was implemented, which BioIR components were reused or extended, how checkpoint tensors were accounted for, and how to construct and run the model. List every check as PASS, FAIL, or PENDING with the command, artifact path, and sample scope. Give the exit code for executed checks and the reason for pending checks. Include documented differences, unsupported features, pending GPU work, and the working directory.
Tips for Model Onboard
- • Start a fresh session after installing so your agent picks up the new skill.
- • This skill ships scripts. Read them before the first run, and run it in a project without secrets first.
- • It is part of a pack with 30 skills. Install only the ones you need with --skill to keep triggers precise.
- • If two skills overlap, mention the one you want by name.
Questions about Model Onboard
What does the Model Onboard skill do?
Implement an upstream pretrained model natively in BioIR using its optimized components, with checkpoint conversion and numerical validation. Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.
How do I install Model Onboard?
Run "npx alyn-cli add model-onboard" in your project. The Alyn CLI installs it for Claude Code, Cursor, Codex and 70+ other agents. In Claude on the web, upload the skill folder as a ZIP under Customize > Skills.
Does Model Onboard work in Cursor and Codex?
Yes. Model Onboard uses the open SKILL.md format, so it works in Claude Code, Claude, Cursor, OpenAI Codex, Gemini CLI, GitHub Copilot and other agents that support agent skills.
When should I use the Model Onboard skill?
Use when adding a new model or architecture to BioIR; includes data pipeline integration when requested.
What is inside the Model Onboard skill?
The SKILL.md covers Establish the Scope, Execution and Evidence, Report, and the skill folder has 17 files including scripts.
What is a good alternative to Model Onboard?
Architecture Zoo by Aperivue is the highest-ranked similar skill from another creator, with an Alyn Score of 44 compared to 78 for Model Onboard.
Is Model Onboard free?
Model Onboard is open source on GitHub under the Apache-2.0 license. It is free to install; check the license before commercial use.
Who made Model Onboard?
Model Onboard is published by NVIDIA in the NVIDIA/skills repository, together with 29 other skills.


