Skip to content

Kermt continue pretrain vs Kermt pretrain scratch

Kermt continue pretrain is an open-source agent skill for health and science. Kermt pretrain scratch is an open-source agent skill for health and science.

  • Kermt continue pretrain is best for continue pretraining a kermt model on my own smiles library.
  • Kermt pretrain scratch is best for train a molecular model on a custom chemistry dataset not covered by released checkpoints.
What it is
Kermt continue pretrain is an open-source agent skill for health and science.
Kermt pretrain scratch is an open-source agent skill for health and science.
Best for
Continue pretraining a KERMT model on my own SMILES library
Train a molecular model on a custom chemistry dataset not covered by released checkpoints
Kind
Skill
Skill
Pricing
Open source
Open source
Platforms
Claude Code, Codex
Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
GitHub stars
4K
4K
Last commit
Nov 2026
Nov 2026
License
Apache-2.0
Apache-2.0
Open source
Good for
Continue pretraining a KERMT model on my own SMILES library; Resume an interrupted molecular pretraining run from its last checkpoint; Check whether a checkpoint and CSV are valid before training
Train a molecular model on a custom chemistry dataset not covered by released checkpoints; Reproduce a KERMT pretraining configuration end to end without a starting checkpoint; Compare vocab, cmim and hybrid pretraining objectives on the same corpus

What Kermt continue pretrain does

  • Checkpoint validation

    Checks that the input checkpoint is a supported pretrain type and redirects users when it is a finetuned or encoder-only file.

  • Vocabulary pass-through

    Reuses the checkpoint's atom, bond and SMILES vocabulary files so the new corpus matches the existing model heads.

  • Containerized training

    Launches the pretraining runner detached inside a container with mounted host directories for data, checkpoints and outputs.

  • GPU batch sizing

    Chooses default batch sizes for single or multiple GPUs and offers a VRAM-based table for smaller hardware.

  • Optional released model fetch

    Downloads the pinned NV-KERMT-70M-v2 bundle over HTTPS only with explicit consent or an opt-in flag.

  • Resumable runs

    Supports fresh-schedule continuation or a true resume of an interrupted run with its saved schedule.

What Kermt pretrain scratch does

  • Fresh vocabulary building

    Builds atom and bond vocabulary directly from the supplied corpus instead of reusing a released one.

  • Explicit objective selection

    Requires choosing vocab, cmim or hybrid pretraining objectives before any run starts.

  • Runtime estimation

    Estimates wall time from corpus size, epochs and GPU count and asks for confirmation first.

  • Detached container runs

    Launches training inside the kermt container in the background for long jobs.

  • GPU and batch sizing

    Suggests batch sizes by GPU memory class and supports selecting specific GPUs.

  • Run manifest and logs

    Writes a reproducible run manifest, training log and TensorBoard output for each run.

Kermt continue pretrain vs Kermt pretrain scratch: questions

Is Kermt continue pretrain or Kermt pretrain scratch cheaper?

Both are open source. Compare current plans on each official site.

What is the difference between Kermt continue pretrain and Kermt pretrain scratch?

Kermt continue pretrain is an open-source agent skill for health and science. Kermt pretrain scratch is an open-source agent skill for health and science.

Keep comparing