Modelpublic

jvonrad/Qwen-2.5-7B-CM-Align-10k

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes7downloads

Model Card

Qwen-2.5-7B-CM-Align-10k

CM-Align (Zhang et al., EMNLP 2025 Findings): English-pivot self-supervised DPO on 10,000 facts.

A LoRA adapter (r=64, alpha=128) over `Qwen/Qwen2.5-7B`, 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 (`Qwen/Qwen2.5-7B`)	51.25	5.35	62.36	26.17	63.55	47.72	35.78
This model	53.89	6.96	63.51	26.35	61.23	53.36	38.99

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B", dtype="bfloat16",
                                            device_map="auto")
model = PeftModel.from_pretrained(base, "jvonrad/Qwen-2.5-7B-CM-Align-10k")
tok = AutoTokenizer.from_pretrained("jvonrad/Qwen-2.5-7B-CM-Align-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}
}