This is a warm-start checkpoint: SFT'ing with IVON yields not just point weights but an approximate Gaussian posterior over them (a mean and a diagonal Hessian/precision estimate). That posterior is the learned prior used to seed the 3PO RLVR runs (B3PO / M3PO / C3PO), where weight perturbations sampled from it drive parameter-space exploration.

Training


Foundation model	`allenai/Olmo-3-1025-7B`
Stage	Warm-start SFT
Data	Llama-Nemotron Post-Training Dataset (SFT subset)
Optimizer	IVON, lr `50.0`, ESS (λ) `1e10`
Hardware	8× NVIDIA H200 (144 GB)

Usage

Loads as a standard causal LM:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("BayesRL/Olmo3-IVON-SFT-7B")
tok = AutoTokenizer.from_pretrained("BayesRL/Olmo3-IVON-SFT-7B")

To use it as the warm-start prior for 3PO RLVR, load the IVON optimizer state via IVON_INIT_METHOD=trained in the companion code's run_rl.sh.

Citation

bibtex

@misc{venkatkrishna2026parameterexploration,
      title={Parameter Exploration for RLVR via Variational Learning}, 
      author={Vatsal Venkatkrishna and Nico Daheim and Iryna Gurevych},
      year={2026},
      eprint={2608.09805},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.09805}, 
}