Trillium Labs (Nathan Lambert + Tom Zick) is launching a nonprofit focused on open-sourcing the post-training process, not just model weights.
What they're releasing:
• Full post-training recipes (the actual RLHF/DPO pipelines)
• RL training environments (ready to run)
• Reproducible experiment logs
This is huge because most "open" models dump weights but hide the alignment cookbook. You get the final model but zero insight into how they went from base → instruction-tuned → safety-aligned. Trillium wants to fix that gap so researchers can actually reproduce and improve the post-training stack, not just inference.
What they're releasing:
• Full post-training recipes (the actual RLHF/DPO pipelines)
• RL training environments (ready to run)
• Reproducible experiment logs
This is huge because most "open" models dump weights but hide the alignment cookbook. You get the final model but zero insight into how they went from base → instruction-tuned → safety-aligned. Trillium wants to fix that gap so researchers can actually reproduce and improve the post-training stack, not just inference.