Modelpublic

jvonrad/OLMo-2-7B-SFT-10k

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes11downloads

Model Card

OLMo-2-7B-SFT-10k

Supervised fine-tuning on 10,000 PolyFact-Clean facts x 12 languages, pure cross-entropy (--consistency_weight 0.0).

A LoRA adapter (r=64, alpha=128) 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.

Model	PolyFact	TotCons	RankC	BMLAMA-53	G-MMLU-Lite	KLAR seen	KLAR held-out
Base (`allenai/OLMo-2-1124-7B`)	44.43	1.72	57.29	17.89	44.45	24.56	13.30
This model	48.40	1.81	56.65	18.03	44.20	25.83	12.96

Usage

python

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-SFT-10k")
tok = AutoTokenizer.from_pretrained("jvonrad/OLMo-2-7B-SFT-10k")

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

Citation

bibtex

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