venus-finetune-workflow

name	venus_finetune_workflow
description	Fine-tune and run custom protein models on VenusFactory (CSV/HF → config → train → predict). Use when the user brings labeled sequences, wants adapter training (ProtT5/ESM2/Ankh/QLoRA notes), or batch inference with a trained config. Do NOT use for zero-shot mutation without labels (zero_shot_mutation_workflow) or built-in function heads already covered by predict_protein_function.
license	Apache-2.0
metadata	{"version":"1.0","skill-author":"VenusFactory2"}

Venus Fine-Tune Workflow

Overview

Orchestrates training hub tools. Expect CSV columns such as aa_seq / sequence + label. GPU recommended; training can be long — set clear success criteria (metrics file / checkpoint exists).

Project Tools (VenusFactory2)

Tool	Args	Purpose
generate_training_config	csv_file or dataset_path; optional valid/test CSV; user_requirements; output_name	Build training JSON
train_protein_model	config_path	Run training; auto-registers under ckpt/user_trained/<model_id>/
register_trained_model	config_path; optional model_id / output_dir / model_path	Explicit re-register (usually unnecessary after train)
list_trained_models	—	List cross-session registered models
protein_model_predict	config_path or model_id + sequence or csv_file	Inference
agent_generated_code	task_description, input_files, output_dir	Splits, metrics plots, CSV cleanup (Expert only; not MCP)

Mode: local only. Online mode disables these tools (and VenusMine / FoldSeek discovery). Science Agent reaches train tools via MCP; Expert via the LangGraph tool hub.

Workflow

1. Validate/split data with agent_generated_code if needed (70/15/15) — Expert path.
2. generate_training_config with explicit user_requirements (model, epochs, LR, QLoRA…).
3. train_protein_model on the returned config path (dependency:step_N:file_path). On success, use registered_model_id / registered_config_path from the result.
4. protein_model_predict with the registered model_id (or config path) for hold-out or new sequences — works across sessions.
5. Figure: loss/accuracy curves from training metrics (dpi≥300).

When NOT to use

• Unlabeled mutagenesis ranking → zero_shot_mutation_workflow
• Off-the-shelf solubility/temp heads → predict_protein_function (protein_property_prediction)

Common mistakes

• Omitting both csv_file and dataset_path
• Predicting without the same config_path used for training
• Writing artifacts outside the session output_dir