forge high level

1. YAML Config

You choose:

• task
• model
• training mode

Example:
MNLI + DistilBERT + SFT

2. forge.py

Main entrypoint

Reads the config
and decides what to run

3. Prepare Run

• load dataset
• load tokenizer/processor
• load model
• choose trainer

4. Train

Possible paths:

• SFT
• KD
• RL

Your first run:
SFT / CE-only

6. Save Outputs

Output dir contains:

• config
• metadata
• logs
• checkpoints
• final model

5. Track Progress

During training:

• logs
• evaluation
• checkpoints
• resume state