Skip to content
mcp/skillhub

Agent Skill

Kermt Pretrain Scratch skill

kermt-pretrain-scratchNVIDIA/skillsskills/bionemo-kermt-pretrain-scratch/SKILL.md

Pretrain a fresh KERMT model from scratch on a user-provided corpus. Builds a new vocabulary from the corpus, instantiates the model architecture from defaults, and launches pretrain_ddp.py inside the kermt container (detached for long runs). Unlike kermt-continue-pretrain, no starting checkpoint is loaded — the model is randomly initialized.

3.7kApache-2.0

always on

82 tokname + description

once used

2,370 tokSKILL.md body

bundled files

116 scripts

repo stars

3.7kGitHub

Install the Kermt Pretrain Scratch skill

Skills CLI (Claude Code, Codex, Cursor and others)

npx skills add https://github.com/NVIDIA/skills --skill kermt-pretrain-scratch

Claude Code, by hand (all projects)

git clone --depth 1 https://github.com/NVIDIA/skills.git /tmp/skills
mkdir -p ~/.claude/skills
cp -R /tmp/skills/skills/bionemo-kermt-pretrain-scratch ~/.claude/skills/kermt-pretrain-scratch

For one project only, copy the folder into .claude/skills/ in that project instead. The agent picks it up when a task matches its description.

What it costs your context

Agent Skills load in three steps. Tokens counted with the o200k tokenizer; your model's count will differ a little.

WhenWhat loadsCost
Every sessionName and description82 tok
When the agent uses itThe SKILL.md instructions2,370 tok
Only if opened11 bundled files (163 KB)varies

Until it is used, kermt-pretrain-scratch costs about 17× less than the median MCP server we measured (1,363 tokens on every turn). Installing all 150 skills from NVIDIA/skills would load 20,796 tokens of descriptions. Skills vs MCP study

Checks

  • Ships scripts the agent can run. Read them before installing.
  • The name doesn't match its folder name, as the specification requires.

Checked 2026-10-22 · method

What's inside

The headings of its SKILL.md. Read the full file on GitHub.

  • kermt-pretrain-scratch
  • Skill and runtime paths
  • Hardware requirements
  • When to invoke
  • Inputs
  • Workflow
  • Hard rules
  • Common errors
  • What's in the manifest after a from-scratch run
  • Replayability

More skills from NVIDIA/skills

All skills in this repo
  1. accelerated-computing-cudfNVIDIA/skills

    Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.

    3.7k2356 tok idle · 2,307 in use
  2. ambient-healthcare-agent-with-nemotron-voice-agentNVIDIA/skills

    Customize NVIDIA Nemotron Voice Agent's Generic Pipecat example for healthcare appointment, five-field patient intake, or custom tool-calling workflows without a separate backend.

    3.7k1744 tok idle · 2,849 in use
  3. amc-run-rtsp-calibrationNVIDIA/skills

    Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.

    3.7k159 tok idle · 4,607 in use
  4. amc-run-sample-calibrationNVIDIA/skills

    Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.

    3.7k164 tok idle · 3,627 in use
  5. amc-run-video-calibrationNVIDIA/skills

    Calibrates pre-recorded `cam_*.mp4` datasets through the AutoMagicCalib REST API. Use for user-supplied local MP4s; route live RTSP streams to `amc-run-rtsp-calibration`.

    3.7k157 tok idle · 4,576 in use
  6. amc-setup-calibration-stackNVIDIA/skills

    Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.

    3.7k67 tok idle · 4,277 in use
  7. boltz2-nimNVIDIA/skills

    Use Boltz2 NIM for biomolecular structure prediction and binding affinity. Invoke for Boltz2, protein structures, protein-ligand/DNA/RNA complexes, SMILES or CCD ligands, pIC50/IC50 affinity scoring, mmCIF output, hosted NVIDIA API calls, or local Docker deployment.

    3.7k71 tok idle · 1,433 in use
  8. cdf-anomalygenNVIDIA/skills

    Use when running the CDF AnomalyGen pipeline over a defect dataset — mask placement, fine-tuning, synthetic defect-image generation (SDG), evaluation, quality refinement, and pseudo-labeling — even when the user names only one stage.

    3.7k54 tok idle · 9,563 in use