cs.CLNov 15, 2026

Ambiguous Examples Help, But Full Data Wins: Revisiting Dataset Cartography for Extractive Adversarial QA

Authors: Paul K. Mandal

Organizations: Neurint LLC, Baton Rouge, LA · U.S. Army Cyber Corps, U.S. Army Reserve, Grand Prairie, TX · Northwestern State University of Louisiana, Natchitoches, LA

Abstract

Dataset cartography argues that training dynamics can reveal regions of a dataset that matter differently for learning: easy-to-learn examples aid optimization, hard-to-learn examples often expose annotation problems, and ambiguous examples contribute strongly to out-of-distribution generalization. A striking implication of the original work is that a model trained on only 33% ambiguous examples can outperform full-data training on several OOD evaluations, including a QNLI reformulation of SQuAD evaluated on Adversarial SQuAD. We revisit this claim in a stricter setting: open-ended extractive question answering, where models must predict answer spans rather than classify question-sentence pairs. For each encoder family, we compute clean-source SQuAD training dynamics, train on full, random, easy-ranked, ambiguous-ranked, and hard-ranked subsets, and evaluate robustness on adversarial-only AddSent and AddOneSent rows. Ambiguous-ranked 33% subsets beat random 33% subsets on exact-match accuracy in all 14 model-evaluation cells and on F1 in 12/14 cells. However, full-data training beats ambiguous-ranked training in 12/14 exact-match cells and 12/14 F1 cells, with mean full-minus-ambiguous gains of 3.40 EM and 4.10 F1. The most thoroughly replicated ELECTRA-small panel is also the main exception to the ambiguous-over-random F1 gains. These results show that clean-source cartography remains useful for reduced-data selection, but its OOD advantage does not carry over as a general replacement for full-data training in extractive adversarial QA.

Figures & tables

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Feb 27, 2026

LFQA-HP-1M: A Large-Scale Human Preference Dataset for Long-Form Question Answering

Long-form question answering (LFQA) demands nuanced evaluation of multi-sentence explanatory responses, yet existing metrics often fail to reflect human judgment. We present LFQA-HP-1M, a large-scale dataset comprising 1.3M human pairwise preference annotations for LFQA. We propose nine rubrics for answer quality evaluation, and show that simple linear models based on these features perform comparably to state-of-the-art LLM evaluators. We further examine transitivity consistency, positional bias, and verbosity biases in LLM evaluators and demonstrate their vulnerability to adversarial perturbations. Overall, this work provides one of the largest public LFQA preference datasets and a rubric-driven framework for transparent and reliable evaluation.
Sep 17, 2026

Intrinsic Sequence-Likelihood Confidence in Retrieval-Dominated Extractive QA: Two Pre-Specified Negatives, and What They Do and Do Not Attribute

In extractive document question answering whose questions were generated from the passages that contain their answers -- so that retrieval recovers 92-99.8% of what any mode combination could reach, whatever its absolute accuracy -- confidence-driven mechanisms have little to gain. Fine-tuning an open language model on a specialized domain corpus yields a model whose own confidence is a tempting control signal: it could decide which queries warrant further adaptation, and which answers to trust. We evaluate both uses under criteria fixed before the runs were executed, across four 7-9B model families whose adaptation moved closed-book F1 by at most +0.03, and both fail: a distillation trigger on all four families, under its pre-specified three-step transfer budget, and a routing-and-abstention policy in its single-model pilot. Retrieval alone recovers 92-99.8% of best-case combined accuracy under every correctness criterion we test, leaving routers no meaningful gain. The sequence-likelihood signal is insufficient relative to that mode -- area under the receiver operating characteristic curve 0.65-0.81 under the registered criterion -- before adaptation as well as after, unchanged by scalar recalibration and not consistently improved by token-level temperature rescaling. And the finer diagnostics depend on the correctness criterion and on answer length; on the three adapted combinations where we could test it, selector ablations show no statistically detectable downstream benefit from the confidence term on any seed; on Gemma, removing it changes the selector from failing to passing both registered criteria. The usable product is a set of pre-specified negatives with their dependencies made explicit.
Jun 23, 2026

OpenThoughts-Agent: Data Recipes for Agentic Models

Agentic language models dramatically expand the applications of AI yet little is publicly known about how to curate training data for broadly capable agents. Existing open efforts such as SWE-Smith, SERA, and Nemotron-Terminal typically target a single benchmark, leaving open the question of how to train models that generalize across diverse agentic tasks. The OpenThoughts-Agent (OT-Agent) project addresses this gap with a fully open data curation pipeline for training agentic models. We conduct more than 100 controlled ablation experiments to systematically investigate each stage of the pipeline, yielding insights on the importance of task sources and diversity. We then assemble a training set of 100K examples from our pipeline and fine-tune Qwen3-32B on this dataset, which yields an average accuracy of 44.8% across seven agentic benchmarks and a 3.9 percentage point improvement over the strongest existing open data agentic model (Nemotron-Terminal-32B, 40.9%). Moreover, our training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons. We publicly release our training sets, data pipeline, experimental data, and models at openthoughts.ai to support future open research on agentic model training.