# Codonfm Finetune

Fine-tune public CodonFM Encodon checkpoints on labeled coding-sequence or...

- Repository: https://github.com/NVIDIA/skills (3812 stars)
- Category: Engineering
- License: Apache-2.0
- Last updated: 2026-10-30
- Runs scripts: yes, review before use

## Install

Claude Code: `npx skills add NVIDIA/skills --skill codonfm-finetune -g -a claude-code -y`
Codex: `npx skills add NVIDIA/skills --skill codonfm-finetune -g -a codex -y`
Cursor: `npx skills add NVIDIA/skills --skill codonfm-finetune -g -a cursor -y`
Gemini CLI: `npx skills add NVIDIA/skills --skill codonfm-finetune -g -a gemini-cli -y`
Claude Code plugin marketplace: `/plugin marketplace add NVIDIA/skills`

## Description

Fine-tune public CodonFM Encodon checkpoints on labeled coding-sequence or coding-variant data using LoRA, head-only, or full fine-tuning. Use when a user explicitly asks to fine-tune CodonFM or Encodon for regression or classification. Support generic public-v1 Encodon workflows only; reject Decodon, MissenseDataset, missense_synom_agg, and generation workflows.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Fine-tune public Encodon

Use `--pretrained_ckpt_path` for public v1. Do not substitute
`--checkpoint_path`: the public runner does not forward that argument to the
fine-tuning task.

## Instructions

Resolve the target label, dataset, checkpoint, and output directory from the
request and available files. Reuse existing data and weights. For training,
check the project's ML dependencies and a compatible NVIDIA GPU before launch.
If a required resource is unavailable, complete the available data preparation
and return the command with the missing prerequisite clearly identified.
When training is requested and the prerequisites are met, execute it and check
the resulting checkpoints and metrics. A request for preparation ends with the
validated inputs and command.

Use the user's labeled dataset when provided. For a demonstration of sequence
regression without a dataset, use the public human RiboNN translation-efficiency
data below and state that choice. This is not a substitute for a user's intended
assay or for labeled coding variants. If variant labels are missing, return the
required schema and a command template promptly; do not search for labels or
invent measured effects.

Default demonstration checkpoint: `nvidia/NV-CodonFM-Encodon-80M-v1`, revision
`399ca9fe17b57941a7bebc6788033919b417413c`, file
`NV-CodonFM-Encodon-80M-v1.safetensors` with sibling `config.json`.
The [public weights](https://huggingface.co/nvidia/NV-CodonFM-Encodon-80M-v1)
are about 307 MB. Download them when needed for the requested work; input
preparation can record an intended checkpoint path. These are the original Encodon weights; the `-TE-`
checkpoints use the separate TransformerEngine implementation.

For an unsupported Decodon or missense-aggregation request, inspect the public
parser/model configuration, explain the missing feature, and finish. Do not
implement the missing model or search private repositories.

## Examples

The rest: https://github.com/NVIDIA/skills/blob/HEAD/skills/bionemo-codonfm-finetune/SKILL.md