feat(trl): improve eval dataset handling and documentation
#3
by
evalstate
- opened
Improve TRL skill documentation and examples to prevent training hangs caused by missing eval datasets.
What's Fixed
- Training jobs hanging when
eval_strategyis set withouteval_dataset - Missing documentation on eval dataset requirements
- Outdated example patterns in training guide
Key Improvements
- Add train/eval split to all training examples (SFT, DPO)
- Add critical troubleshooting section with correct/wrong patterns
- Show examples of both approaches (with eval / without eval)
- Update references to point to production examples
- Make utility scripts executable
thank you haiku, you did an amazing job....
evalstate
changed discussion status to
closed