01.First Run Explained

First run command

python forge.py --config configs/classification/mnli/distilbert/00_sft.yaml --max_steps 10

Why this is the first run

• It is a short smoke test.
• It verifies that your environment, config loading, data pipeline, and training loop all work together.
• It uses a small step budget (--max_steps 10) so failures appear quickly.

Command breakdown

• python forge.py
  • Runs the main Forge training entrypoint.
• --config configs/classification/mnli/distilbert/00_sft.yaml
  • Selects the experiment config:
    • model family: DistilBERT model
    • task: MNLI
    • paradigm: SFT (Supervised Fine-Tuning)
• --max_steps 10
  • Limits training to 10 steps for fast validation.

What should happen if it is healthy

• Config is parsed successfully.
• Dataset and tokenizer load without errors.
• Model initializes correctly.
• Training starts and runs for 10 steps.
• Metrics/logs are emitted.
• Process exits cleanly without runtime errors.

Common failure points

• Missing environment variables in .env.
• Conda environment not activated.
• Dependency mismatch or missing package.
• Wrong config path.
• Data path or dataset access issues.

What to do after this passes

1. Remove or increase --max_steps.
2. Run the full experiment from the same config.
3. Monitor metrics and checkpoint outputs.