Hi,

I was in touch with @jacob-morrison a few weeks ago when I was trying to reproduce the OLMo-3 SFT checkpoints.

I've been working on finetuning OLMo-3 7B from base to instruct using OLMo-core + open-instruct without the Beaker infra dependency, and I've put together some documented workflows that might be useful for others in the community. I wanted to check if any of this would be a good fit for a PR. I was able to reproduce the performance of the 7b-instruct-sft checkpoint.

What I have:

1. End-to-end SLURM training workflows

• Single-node training scripts for Think SFT and Instruct SFT stages
• Data preparation scripts that convert HuggingFace datasets to OLMo-core tokenized format

2. Jupyter notebook workflows documenting the complete process:

• Think SFT workflow - from cloning OLMo-core to launching training
• Instruct SFT workflow - second stage finetuning

3. Evaluation pipeline

• vLLM model serving with Gradio chat UI
• LLM-judge evaluation using OpenJury
• Results analysis and visualization scripts

4. Model testing utilities

• Checkpoint conversion (OLMo-core to HuggingFace)
• Test scripts for verifying trained models

Everything is documented with READMEs explaining the setup and usage (see MODEL_TESTING_GUIDE.md). The main value I see is having a reproducible, well-documented path for SLURM users to finetune OLMo-3 without needing Beaker/cloud infrastructure.

Would any of this be useful for open-instruct? Happy to clean things up and submit a PR if there's interest, or just keep it as a reference if it's not a fit.

Thanks!