Probing Factual Knowledge Transfer with Training Data Interventions
Romina Oji, Marc Braun, Marcel Bollmann, Marco Kuhlmann, Jenny Kunz
cs.CL · cs.AI
Quick verdict
论文声称通过训练数据干预，英→波（斯）语的事实迁移非常有限；因果控制良好，但语言对单一、模型规模仅为 5.72 亿（572M），且缺乏统计学严谨性。
Summary
该论文探讨了多语言模型在持续预训练过程中是否能真正跨语言迁移事实知识，还是仅仅回忆从目标语言数据中学到的事实。作者在波斯语 FineWeb-2 上对英语基础模型（基于 SmolLM2，5.72 亿参数）进行了持续预训练，并以三种粒度系统地移除了 500 个 SIFT 三元组（20 种关系，分为“通用”和“波斯相关”来源）——包括共现移除、完整主体移除和中性模板替换——此外还增加了一个随机移除对照组。评估采用了原生编写的波斯语完形填空（cloze）模板，对比了困难（语义上接近）与随机候选集，并得出了迁移、非迁移和排序相关指标。在完整主体移除和困难候选集的条件下，波斯语准确率降至 32.4%，Overall 降至 20.6%，Transfer 降至 39.9%，而经验下限为 26.0%，这表明跨语言迁移是有限的；随机移除对照组（47.2%）表明移除数量并不能解释这一下降，且句子级共现移除留下了可恢复的事实信号。随机候选集夸大了明显的迁移效果（CPT-SubjRM 为 49.9% 对比 39.9%），且迁移随英语端主体频率的强烈变化：波斯相关主体（中位频率为 284，而通用主体为 196K）迁移概率接近于零。局限性：单一语言对、一个 5.72 亿参数的模型系列、没有正式的显著性检验，以及将困难候选集等同于阻断浅层线索干预的未经验证假设。
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under the strictest intervention (full subject removal from Persian training data with hard distractors), cross-lingual fact transfer is limited: Transfer reaches 39.9% against a 26.0% empirical floor, with Overall at 20.6% (Table 2).
- Sentence-level co-occurrence removal does not eliminate fact signal from the target corpus: CPT-CoocRM shows higher Transfer (48.9%) and Non-Transfer (30.1%) than full subject removal (39.9% / 24.4%) on hard candidates.
- Randomly sampled candidate sets substantially inflate apparent transfer by rewarding shallow associative heuristics: CPT-SubjRM Transfer is 49.9% with random vs 39.9% with hard candidates.
- Cross-lingual transfer is gated by source-language entity frequency: Persian-related subjects with median English frequency 284 (vs 196K for general subjects) transfer at near-chance levels.
- Random removal of a comparable text volume (7.79% of Persian tokens) leaves accuracy essentially unchanged (47.2% vs CPT-Full's 46.6% on hard), attributing the drop under subject removal to the removed factual evidence rather than corpus shrinkage.
Strengths
Causal intervention design with a volume-matched control
Three removal granularities plus CPT-RandRM (7.79% Persian tokens removed vs 8.50% for CPT-SubjRM, Table 1) cleanly separate fact removal from corpus shrinkage: accuracy stays at 47.2% under random removal vs 32.4% under subject removal on hard candidates (§5.1), so the drop is attributable to removed fact signal, not data volume.
Dual-purpose resource with native templates
SIFT (§3) supplies triples with exhaustive Persian surface-form variants for complete removal, natively written (not machine-translated) Persian cloze templates (§3.2), per-subject corpus frequency statistics, and a provenance stratification (386 general vs 114 Persian-related triples) absent from LAMA-style and ECLeKTic resources.
Quantified heuristic inflation
Robustness and transparency of reporting
Isolating source frequency from direct exposure
Weaknesses
No formal uncertainty quantification
Approximate causal identification with unverified assumptions
Narrow external validity
Annotation transparency
Format-sensitive ablation conclusion
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: Probing Factual Knowledge Transfer with Training Data Interventions
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the SIFT dataset?
SIFT (Stratified Intervention Facts for Transfer) is a set of 500 English–Persian fact triples across 20 relations, split into 'general' facts with globally prominent subjects (386 triples) and 'Persian-related' facts with subjects rooted in the Persian-speaking world (114 triples). Each entry carries Persian surface-form variants, natively written Persian cloze templates, and hard/random distractor candidates, and is designed both for removing facts from training corpora and for evaluating factual recall. It is released on Hugging Face.
Do language models transfer facts learned in English to other languages?
In this paper's controlled setup — continued pretraining of a 572M English model on Persian data from which the tested facts were deleted — very little transfer was observed: with all subject mentions removed from Persian and hard distractors used, Transfer accuracy was 39.9% versus an empirical floor of 26.0%, and only 20.6% of facts were answered correctly in both languages, meaning most English-acquired facts were not retrievable in Persian.
Why do random candidate sets overestimate factual knowledge in language models?
Randomly sampled distractors are usually semantically distant from the correct answer, so shallow cues (e.g., a Persian-looking name suggesting an Iranian city) suffice to pick the right option without knowing the fact. The paper measures this inflation: the CPT-SubjRM model's Transfer score drops from 49.9% with random candidates to 39.9% with hard, semantically close distractors, and CPT-Full drops from 70.7% to 59.7%.
Is the paper's code and data available?
Yes, according to the paper: the SIFT dataset is on Hugging Face, the intervention and evaluation code is at github.com/liu-nlp/SIFT, and the seven trained 572M-parameter models are stated to be available on Hugging Face.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Right-to-left Persian strings in Figure 1 are extracted in scrambled token order (a PDF artifact); the intervention design remained inferable from the surrounding English text.
- Per-relation values in Figure 5 and frequency-bin values in Figure 3 are extracted as unlabeled digit sequences; all numbers cited in this review were taken from the prose and Tables 1–7, not parsed from figure cells.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.