OLMo-2-7B-GRPO-ladder-s6500

OLMo-2-7B-GRPO-ladder-s6500

Consistency-driven GRPO with a laddered all-correct bonus (--bonus_shape ladder), checkpoint 6,500 of a 10,000-step run.

A LoRA adapter (r=128, alpha=256) over allenai/OLMo-2-1124-7B, trained for the paper Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning. It is one arm of a controlled comparison in which SFT, DCO, CM-Align and GRPO all see the same 10,000 facts from jvonrad/PolyFact-Clean across the same 12 languages, so the methods differ only in objective.

Evaluation

Accuracy (%) unless noted. PolyFact-Clean is the 2,039-fact curated test split with byte-normalised log-likelihood scoring; TotCons is the fraction of facts answered correctly in all 12 languages; RankC is RankC@4 (floor 9.02, chance 37.68). KLAR is free-form generation over 17 languages, split into the 7 seen in training and the 10 held out.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-7B", dtype="bfloat16",
                                            device_map="auto")
model = PeftModel.from_pretrained(base, "jvonrad/OLMo-2-7B-GRPO-ladder-s6500")
tok = AutoTokenizer.from_pretrained("jvonrad/OLMo-2-7B-GRPO-ladder-s6500")

Evaluation used the closed-book prompt Question: {q}\nAnswer: with the options hidden, matching evaluate/evaluate_crosslingual_consistency.py.

Citation

@misc{polyfact2026,
  title  = {Improving Cross-Lingual Factual Recall via Consistency-Driven
            Reinforcement Learning},
  author = {von Rad, Jonathan},
  year   = {2026},
  eprint = {2606.06586},
  archivePrefix = {arXiv}
}