Open Instruct for reproducing post-training experiments

How it works

The repository contains scripts and configurations for post-training workflows tied to research releases. It is most useful when the dataset, checkpoint and recipe can be matched to one experiment.

What you need

Identify the specific paper result, data and checkpoints first. A related training script may be intended for a different model or experiment.

What to try first

Pick one experiment and trace the referenced data, model and evaluation code. Run a small configuration first to verify preprocessing and artifact output.

Check whether the released benchmark setup uses the same templates and model variant as training. Differences can matter more than a small change in hyperparameters.

Where to look for failures

Replicating a paper result requires matching more than the trainer. Track exact data versions and hardware assumptions and verify all artifact terms.

Related Franklin pages

Choose a training objective separately from how you update or store weights. Verify the data format, then reload the resulting artifact against its original base model.

• Train transformer models with documented fine-tuning and preference-learning trainers. TRL for supervised and preference-based training
• Adapt a supported model by training a subset of parameters and saving compact artifacts. PEFT for training and loading model adapters

For the distinction between training from scratch and adapting a checkpoint, see Franklin’s glossary. Fine-tuning

Project documentation

• Current project README
• README at the reviewed revision