Skip to content

Clean up: Clarify Tinker epoch-based training configuration#20

Merged
Sudhendra merged 1 commit intomainfrom
fix/tinker-config-steps-docs
Feb 13, 2026
Merged

Clean up: Clarify Tinker epoch-based training configuration#20
Sudhendra merged 1 commit intomainfrom
fix/tinker-config-steps-docs

Conversation

@Sudhendra
Copy link
Owner

Summary

  • remove the stale cloud.training.steps entry from configs/training.yaml and clarify that Tinker run length is epoch-driven in the current training script
  • update active docs examples (docs/AGENTS.md, docs/SETUP.md) to stop showing the obsolete steps=300 argument
  • keep runtime behavior unchanged while eliminating a misleading config/docs mismatch before v2 cloud training

Test Plan

  • .venv/bin/python scripts/train_tinker.py --estimate-cost --config configs/training.yaml
  • .venv/bin/python -c "from pathlib import Path; from src.training.train_tinker import load_config_from_yaml; c=load_config_from_yaml(Path('configs/training.yaml')); print({'model': c.model, 'epochs': c.epochs, 'batch_size': c.batch_size, 'learning_rate': c.learning_rate})"
  • .venv/bin/python - <<'PY'
    • parse all data/training/{train,valid,test}.jsonl records and verify max approximate token length is below 2048 (result: max 1273)

@Sudhendra Sudhendra changed the title Clarify Tinker epoch-based training configuration Clean up: Clarify Tinker epoch-based training configuration Feb 13, 2026
@Sudhendra Sudhendra merged commit 1b23fd8 into main Feb 13, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant