Language models are commonly discussed as technical artefacts, but they are obviously shaped by the linguistic worlds conveyed by data during their training. Using Italian language models as evidence, I want to bring attention to the nature of the systems which result from training and specialising models on translated and synthetic data, and further curating them, and to the meaning of testing them on equally unnatural data. Are these eventually models of Italian? Are they models of language? Does NLP still care about language? These questions yield another, more concrete question: what language do we actually want language models to produce? I argue that this question cannot be answered if we do not first consider a clearer distinction between language models designed as technical products and language models designed as tools for studying language itself. The answers then might be diverse, the languages we are talking about might be diverse, and the picture might not be as pessimistic as we fear.
Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.
The online culinary ecosystem is increasingly populated by recipe content generated, modified, or summarized by Large Language Models (LLMs). While often plausible, such outputs may contain hallucinated ingredients, misrepresented quantities, or culturally implausible combinations, limiting their suitability for downstream applications and knowledge graph construction. In this paper, we present a semi-automated soundness assessment workflow for validating structured recipe data extracted and augmented by LLMs from informal culinary sources. Developed as part of FKG(.in), a knowledge graph of Indian food, the pipeline identifies and addresses common failure modes, including structural inconsistencies, semantic and logical incoherence, and deviations from the source text, through a multi-stage process combining formal grammars, vocabulary-based checks, statistical heuristics, Set Transformer-based coherence modeling, and retrieval-based verification. Although evaluated on Indian recipes, the proposed methods are applicable to broader multilingual and multicultural culinary domains. We provide a practical, auditable, and application-agnostic framework for validating LLM-augmented recipe data, thereby strengthening the foundations of machine-readable food knowledge infrastructures in the era of LLM-generated content.
Preference learning optimizes models using response pairs, yet the informativeness of these pairs is fundamentally shaped by the instructions from which they are generated. We identify instruction quality as a hidden bottleneck in preference learning: low-quality or ambiguous instructions restrict the response-quality distribution, limiting strong chosen responses and weakening preference signals. Through Best- and Worst-of-N analyses, we show that instruction quality constrains both the ceiling and floor of sampled response quality. Motivated by this observation, we introduce an instruction-refinement pipeline that selects weak instructions using reward signals and revises them with rubric-guided LLM feedback, improving preference data without discarding examples. Across offline and online preference learning settings, experiments on multiple models and benchmarks show broad alignment improvements over original data and alternative data-improvement strategies. Further analyses indicate that instruction refinement raises achievable response quality and complements response-centric preference data curation. Overall, instruction quality emerges as a key factor governing how informative preference signals are formed for LLM alignment. Code is available at: https://github.com/01choco/instruction-refinement/
PDF corpora advertise their size in tokens but compute every rate they publish (coverage, OCR routing, re-fetch recovery, language mix) per document, and none decomposes its token total. The two units diverge sharply. On CC-MAIN-2021-31-PDF-UNTRUNCATED (7.9M web PDFs, 32.6B tokens), 3.02% of text-bearing documents hold half the tokens (Gini 0.807); documents over 50 pages are 5.00% of the corpus but 53.53% of its text. The PDFs produced by a TeX{} toolchain are 1.66% of documents and 4.05% of the text. The clearest casualty is Common Crawl's truncation cap: it affected 23.06% of documents and 63.08% of the text. Reconstructing the truncated files and extracting both versions, two widely used libraries recover 11.4% and 1.4% of that text; between 72% and 97% of affected documents yield nothing; roughly 55--62% of the corpus's text is lost. Under the 5 MiB cap adopted in March 2025, 30.19% of tokens would still be truncated, and recovery on those documents rises only from 3.3% to 13.2%. We recommend that corpus statistics be reported in both units: documents and tokens.
LLMs have been increasingly used to catch data quality issues automatically, but we know very little about how consistent these judgments actually are. This study tests an LLM on two e-commerce data quality tasks, entity matching and brand mislabeling, against rule based baselines and human verified ground truth, under both zero-shot and few-shot prompting. On entity matching while using the Abt Buy benchmark (2,194 labeled pairs), a simple rule based baseline (F1=0.950) performed about as well as LLM zero shot prompting (F1=0.948). Moreover, a few-shot prompt revision that looked effective on a small validation sample reduced full-scale performance to F1=0.914. This showed that small sample prompt evaluation can be misleading. On brand mislabeling detection, using 500 Amazon product listings with synthetically injected labeling errors, the LLM clearly outperformed a naive rule based baseline (F1=0.833 vs 0.721), because it could draw on background knowledge of brand product relationships that a simple rule could not access. Testing consistency across repeated runs (200 pairs, 5 runs at temperature 0.7) showed the model agreeing with itself 99.7% of the time on average, with 99% of pairs giving identical answers across all 5 runs. Using majority voting across these runs only improved F1 by 0.005, at 5 times the inference cost. These results suggest that the value of using an LLM over traditional methods depends heavily on the task. LLMs offer little advantage when strong lexical signals already exist, but a clear advantage when the task requires background knowledge, all while remaining highly consistent across repeated queries.
We study baking documents directly into the weights of a 4-bit Gemma-4-e4b model via LoRA, so a system can answer questions about a corpus closed-book: no retrieval and no context-window budget. Across roughly 100 training runs from single documents to a 99-document corpus, we find that once adapter capacity is adequate, training-data quality is the dominant lever on closed-book accuracy, outweighing LoRA rank, learning rate, and two alternative architectures combined; capacity itself is a hard gate below which no data intervention helps. A single curation pass (shortening gold answers to canonical 1-6 word spans and dropping trivia) moved closed-book accuracy from 57.7% to 85.7% on a 15-document corpus, a larger jump than any architectural change. We confirm a capacity trend (rank must grow with corpus size) entangled with a coupling between rank and learning rate that we initially misdiagnosed. On a 15-document slice we add a real retrieval baseline: the internalized adapter (84.2% recall) beats a BM25-RAG pipeline with a base reader (58.9%) and even a realistic gold-chunk oracle (65.6%) at lower latency. We report the full arc, including three misdiagnoses, as a case study in debugging LLM training empirically.
Jungseob Lee, Seungyoon Lee, Suhyune Son +4cs.CL cs.AI
A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.
Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.
As language modeling technology matures, there is an increasing research focus on the composition and curation of datasets used to train these models. For instance, practitioners commonly seek to augment high-quality datasets with additional text to enhance the performance of models trained on that data. However, informed decisions about data augmentation require more nuanced assessments about data quality. We build on work measuring the precision and recall of generative models to develop a pair of metrics that quantify (1) fidelity, capturing how closely candidate text resembles reference data, and (2) diversity, capturing how well it covers the modes of the reference dataset. Our metrics are based on optimal transport divergence functionals between discrete text summaries. In experiments on M2D2 text datasets, we show that these metrics are able to disentangle a lack of fidelity from a lack of diversity in deficient candidate text. In further experiments, our metrics detect diversity deficits in synthetic GSM8K-style math datasets, which correlate with degradations in downstream accuracy of language models finetuned on this synthetic data.
The continuous evolution of large language models drives escalating demands on data scale and quality, and as different training stages impose increasingly tailored data requirements, systematic organization of high-quality corpora becomes indispensable. Existing corpus construction pipelines confine the resulting corpora to flat, undifferentiated document collections, universally lacking systematic knowledge organization. We present Cortex, to our knowledge the first framework that elevates web-scale corpus construction from flat document filtering to structured knowledge organization through an Ontological Corpus Graph (OCG), a three-layer heterogeneous structure unifying a quality-refined content layer, a hierarchical lightweight ontology layer via LLM-driven automated evolution, and a cross-domain alignment layer enabling inter-domain association at arbitrary taxonomic resolution. Comprehensive experiments confirm the effectiveness of Cortex. In particular, we leverage the OCG to synthesize CortexBench, a cross-domain search-and-reasoning benchmark whose evaluation across eight frontier LLMs validates the effectiveness of quality refinement, domain organization, and cross-domain data synthesis. We will publicly release the complete codebase, a 24.14B-token refined corpus with its OCG, and CortexBench.
Several of the world's languages are still under-resourced in terms of Natural Language Processing (NLP) tools. This is mostly due to the lack of high-quality datasets to train, develop, and evaluate systems and models for several tasks, such as Machine Translation (MT). We conduct a manual audit of the parallel and monolingual corpora available for Lombard, an under-resourced language continuum from Italy. Our analysis reveals that the perceived abundance of web-scraped data is an illusion, with massive datasets plagued by severe language misidentification, boilerplate text, and non-linguistic noise. Furthermore, we analyze the orthographic composition of the valid Lombard portions across web-scraped datasets, curated corpora, and benchmarks. Our findings show conflicting orthographical systems and severe representational bias across all corpora: high-quality data is heavily skewed towards Western Lombard varieties, with Eastern ones left on the margins. This underscores the need for variety-aware, community-driven data curation rather than purely quantity-driven scraping.
The widespread use of Large Language Models (LLMs) as writing tools challenges the validity of crowdsourced data, as crowdworkers may outsource tasks to models. To better understand how this is addressed, we surveyed 155 researchers in NLP and related disciplines about their experiences and opinions on collecting free-text responses via crowdsourcing. This paper provides an overview of practitioners' challenges, mitigation strategies, and the foreseen implications on data quality. 44% of respondents reported observing LLM usage in their crowdsourced data. While 93% of them had anticipated this, half were unsure what precautions to take. The most prevalent detection strategies are distinctive textual style patterns and unusually fast completion times. Overall, survey responses show that the research community is aware of the problem and taking measures, but existing efforts remain insufficient to fully address it. Finally, we derive a set of considerations to guide future crowdsourced free-text data collection in the era of LLMs.
Saeid Asgari Taghanaki, Rakshanda Agarwal, Bruce Sun +10cs.LG
Fine-tuning large language models (LLMs) for domain-specific tasks requires training datasets that comprehensively cover the target capabilities a practitioner needs. Yet identifying which capabilities a dataset fails to support, and doing so before an expensive fine-tuning run, remains a largely unsolved problem. We introduce GoalCover, a framework that helps practitioners systematically detect capability gaps in fine-tuning datasets through interactive goal decomposition and automated coverage assessment. GoalCover guides a practitioner through structured decomposition of a high-level goal into atomic, independently evaluable subgoals; assigns each training sample an LLM-based alignment score against every subgoal; and surfaces missing capabilities through automated analysis of low-scoring sample explanations. We validate the framework along two complementary axes. First, through controlled corruption experiments across three domains (medical QA, legal summarization, code generation), we show that GoalCover reliably distinguishes targeted from non-targeted capability impacts: target subgoals degrade by 25.6% on average versus 2.1% for non-target subgoals (Cohen's d=1.24). Second, we demonstrate downstream utility on a financial-summarization Reinforcement Fine-Tuning (RFT) task with Qwen-3-14B: training on GoalCover-filtered data improves the LLM-judge reward from 3.77 to 4.12 (out of 5) over the unfiltered baseline, and combining filtered data with goal-conditioned synthetic samples yields the strongest result (4.20). The two results together show that GoalCover works as a practical pre-fine-tuning diagnostic: it detects capability gaps and produces concrete signal for closing them.