Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A single prompted LLM can rewrite datasets containing private or toxic content into training-ready data, outperforming a commercial redaction service while preserving usefulness.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-04 20:15 UTC pith:NK7FOLEI

load-bearing objection A practical, well-framed method with one genuinely strong large-scale result, but the flagship anonymization claim rests on a distribution-matched synthetic benchmark and needs independent evaluation. the 4 major comments →

arxiv 2509.08653 v2 pith:NK7FOLEI submitted 2025-09-10 cs.LG cs.CL

Generative Data Refinement: Just Ask for Better Data

classification cs.LG cs.CL
keywords generative data refinementgrounded synthetic datadata anonymizationPII removaldetoxificationsynthetic data diversitydata scalingLLM prompting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that a pretrained LLM, given a plain-language prompt, can act as a data refinery, rewriting each sample of a raw dataset so that sensitive identifiers and toxic statements disappear while the otherwise useful content survives. The authors call this Generative Data Refinement (GDR) and position it as a way to unlock the large stock of user-generated data that is currently unusable for training because it risks leaking private information or teaching models to produce harmful text. Their experiments show GDR removes personally identifiable information with far higher recall than a commercial detector service, lowers the measured toxicity of 4chan conversations below even a synthetic chat baseline, and produces refined data that still trains models to answer questions about public facts while giving no correct answers on private ones. Because each rewrite is anchored to a real example, the refined dataset keeps the diversity of the original web-scale data, sidestepping the mode collapse common in purely synthetic generation.

Core claim

Generative Data Refinement (GDR) reframes synthetic data generation as editing: each real data point is passed to a pretrained LLM, which returns a rewrite that satisfies a target criterion — no private identifiers, no toxic content — while staying as close to the original as possible. The central result is that one zero-shot prompt achieves 0.99 recall across 108 PII categories versus the commercial detector's 0.53, and refined 4chan conversations score less toxic than synthetic chat. The refined dataset, the authors argue, inherits the diversity of the real corpus, so models trained on it still learn public facts and never recite private ones.

What carries the argument

The load-bearing mechanism is conditioning: the generative process g(·|x_i) anchors each rewrite to the real sample rather than sampling freely. Concretely, a pretrained LLM is invoked with a shared prompt that instructs it to identify the undesirable content, replace it with a generic placeholder of matching length, and leave everything else unchanged. This 'leave the rest unchanged' instruction distinguishes GDR from synthetic data generation: it preserves diversity, keeps factual numbers intact, and turns the LLM from an open-ended generator into a targeted editor. Few-shot prompting and fine-tuning let a much smaller model match the larger refiner's precision and recall.

Load-bearing premise

The headline claim that GDR outperforms an industry PII service rests on a synthetic benchmark whose sentences were generated by the same model family that does the rewriting, and half of those sentences explicitly name the type of identifier present; real-world documents rarely cooperate that way.

What would settle it

Build an evaluation set from naturally occurring documents with verifiable private content — real leaked API keys, genuine email or chat logs with known identifiers — and ask the GDR prompt to refine them, checking string-for-string whether every original sensitive substring survives. If recall on such real documents falls toward the 0.53 of the commercial detector, or if any original identifier remains recoverable in the refined text, the benchmark-driven claim of 0.99 recall is confined to template-like, model-generated sentences.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Datasets that are currently discarded because they contain PII or toxic content could be kept for training after a rewrite pass, easing the projected shortage of public training data without waiting for new web content to be indexed.
  • Refinement cost can be driven down: a small model fine-tuned on a few thousand examples matches or beats the large refiner's precision and recall, making GDR practical for million-line code corpora.
  • Because GDR keeps a distance measure between the original and rewritten samples, a user can audit the refinement by checking whether the specific sensitive string survived in any output.
  • The same framework transfers to any criterion a language model can judge, since the problem setup treats the target criterion as a plug-in indicator function.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending GDR's audit beyond short QA pairs — checking whether long-form refined outputs still permit inferring or paraphrasing the original private value — would strengthen its guarantees, since the paper's utility tests only cover brief question-answer strings.
  • The detoxified 4chan data made a fine-tuned model markedly harder to identify as AI-generated (31% evasion), hinting that GDR-style rewriting transfers stylistic traits of the source corpus; this could be studied as a deliberate style-transfer tool rather than a side effect.
  • Because the refiner must first recognise sensitive content to remove it, weaker models are likely to hit capability cliffs — reproducing PII instead of rewriting it — so practical deployment would need per-model verification rather than trusting a benchmark score earned by a larger model.
  • Large-scale rewriting changes the token distribution of a corpus (placeholders, paraphrases, normalised toxicity); training on very large GDR-refined volumes could therefore introduce systematic biases not visible in the paper's small-scale fine-tuning tests.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Generative Data Refinement (GDR), a framework in which a pretrained generative model rewrites each example of a real dataset according to a semantic constraint, with the goal of removing undesirable content (PII, toxicity) while preserving useful information for training. The main experiments evaluate GDR as a PII anonymizer against a commercial detector-based service (DIRS) on a synthetic 108-category benchmark and on 1.2M lines of real open-source code with expert line-level annotations; as a detoxifier of 100k 4chan /pol/ messages, with a fine-tuning experiment that uses a quiz extracted from the same data; and as a source of diverse grounded synthetic data, comparing embedding and overlap diversity against prompted synthetic conversations. The paper also studies model-size effects and adaptation via few-shot prompting and supervised fine-tuning.

Significance. The core idea—grounded synthetic data via refinement of real examples rather than sampling from a prompt—is simple and attractive, and the paper's framing around data exhaustion and unindexed user content motivates it well. If the empirical claims hold, GDR would be a practical, low-overhead alternative to removing PII-containing or toxic data from training sets. The paper is honest about failure modes (Section 5 and Appendix M), publishes the prompts, and includes a large-scale real-code evaluation, which are genuine strengths. However, the headline PII benchmark is synthetically generated by the same model family as the refiner, and the detoxification utility evaluation is partly circular; these need additional external-validation experiments before the broad claims in the abstract are sufficiently supported.

major comments (4)
  1. [§4.1, Appendices E–G] The PII benchmark is not an independent test of PII removal. Both PII-positive and PII-negative sentences are generated by Gemini Pro 1.5, the same model family used as the GDR refiner; the positive-generation prompts instruct that the presence of PII be 'obvious' and half of the templates name the PII type (e.g., 'US_PASSPORT_NUMBER'), while the negative-generation prompt enumerates exactly the non-PII numeric categories (scientific measurement, census result, etc.) that the GDR prompt in Appendix A tells the model not to rewrite. GDR's 0.99 recall may therefore reflect pattern-matching to generator cues rather than context-aware detection on arbitrary user text. Because the abstract's flagship claim ('outperform industry-grade solutions for dataset anonymization') rests on this benchmark, the comparison should be rerun on a real-world PII corpus (e.g., Enron emails or expert-annotated
  2. [§4.1, Table 1] The precision and F-score aggregates are computed over different item sets. Precision is only measured for numeric PII categories ('For numeric PII categories, we also compare GDR and DIRS in terms of precision'), while recall is measured over all 108 categories. Table 1 presents a single 'F-score' (0.88) that is not a legitimate harmonic mean of the reported recall and precision over a common support. The 0.80 precision also says nothing about false-positive rates for names, addresses, and other non-numeric PII. Either report precision over all categories or clearly relabel Table 1 as a composite of a macro-recall over 108 categories and a precision over numeric categories only.
  3. [§6.2, Table 4] The pol5k-quiz evaluation is circular as reported. The quiz is generated by the same GDR fact-extraction prompt from pol100k, and the model is fine-tuned on the GDR-refined pol100k; no sentence in §6.2 states that the 5k quiz items were excluded from the fine-tuning set. If they are not excluded, the 0.88→0.92 gain could reflect memorizing the exact QA pairs or distribution matching to the GDR output format, not preservation of world knowledge. Please specify the split, or better, evaluate on an independently constructed question set (e.g., human-written questions about facts in the original pol100k).
  4. [Tables 1–5, Figures 2–5] All comparisons are point estimates with no confidence intervals or significance tests. This is not merely a stylistic request: Table 3 reports a mean toxicity reduction from 0.19 to 0.13 and Table 4 reports an accuracy gain of 0.88 to 0.92—differences that need error bars to support the text's 'significantly' language. Please add bootstrap confidence intervals or equivalent over examples/datasets, especially for the small effect sizes.
minor comments (5)
  1. [§2.2] The paragraph describing DP-SGD and its compute/sample-efficiency trade-offs appears twice verbatim. Remove the duplicate.
  2. [§4.1, §6.2] Figure/table references are mismatched: §4.1 says 'In Figure 1, we report...' but the result appears in Table 1, and §6.2 says 'reported in Figure 4' where the accuracy numbers are in Table 4. Please audit all cross-references.
  3. [Appendix G] Example 2 in the PII-negative template prompt reads 'There are value cells...' instead of 'There are {value} cells...', and there are small typos such as 'sculpter' and an unmatched quote in the JSON format description. Fix these for reproducibility.
  4. [Tables 3 and 5, Figure 5] The aggregation of Perspective API scores is unclear: Figure 5 shows per-category 'Confidence' values while Table 3 reports a single mean. Please state whether the mean is over categories, over examples, or over both, and whether toxicity scores are max-aggregated per example.
  5. [§4, §5] The commercial service 'DIRS' is never expanded or documented beyond 'a commercial method for PII detection'. Since the comparison is a key empirical anchor, please provide the service name/version or enough implementation detail that the comparison is reproducible and interpretable.

Circularity Check

3 steps flagged

Partial circularity: key PII and utility evaluations are generated by the same model/process being tested, but independent code and toxicity experiments support the central GDR claims.

specific steps
  1. other [Section 4.1]
    "Both PII-positive and PII-negative sentences were generated by sampling Gemini Pro 1.5 ... via the prompts in Appendix E – G. ... GDR, based on Gemini Pro 1.5 with a single, shared zero-shot prompt for PII removal across all 108 categories, achieves significantly higher performance than DIRS’s specialized collection of detectors."

    The PII benchmark is generated by the same model family (Gemini Pro 1.5) that serves as the GDR refiner. The generation prompts explicitly instruct that the {pii} variable should be 'obvious' and should be mentioned as 'identifier', 'ID', 'secret', 'account info', or even the PII type name (Appendices E–F). Negative examples are generated with instructions to use the value as a 'scientific measurement, census result' (Appendix G), which Appendix A of the GDR prompt explicitly lists as non-PII. Thus the benchmark's positive/negative distinction is encoded in the very cues that the GDR prompt tells the model to rely on. The high recall is largely a measure of self-consistency with the model's own generation/instruction distribution, not an independent evaluation against real-world PII text.

  2. fitted input called prediction [Section 4.3]
    "We perform standard supervised fine-tuning (SFT) of Flash 8B on a dataset of 10k procedurally-generated PII-positive sentences reserved for training, following a similar protocol to that described for producing the evaluation examples in the PII benchmark. Our results in Figure 2 show that this process significantly improves Flash 8B’s recall and precision on the PII benchmark, allowing it to surpass that of Gemini Pro 1.5."

    The SFT training set is produced by the same procedural generation protocol as the evaluation benchmark, as the paper itself states ('following a similar protocol to that described for producing the evaluation examples'). Consequently, the reported recall/precision improvement on the PII benchmark is an in-sample evaluation on the training distribution, not a prediction of performance on independently sampled real-world PII. The conclusion that small models can be adapted for GDR is fitted to the benchmark generator rather than demonstrated by an external test.

  3. self definitional [Section 6.2]
    "Our detoxification prompt additionally instructs the LLM to extract any facts about the world present in message pairs in pol100k and to reformulate each of these facts into a question-answer pair (see Appendix 10). We thus produce pol5k-quiz, a dataset of 5k subsampled question-answer pairs, whose requisite knowledge is present in pol100k. ... We use pol5k-quiz to measure to what degree GDR’s detoxified outputs preserve otherwise non-toxic content in pol100k."

    The facts and QA pairs used as ground truth in pol5k-quiz are produced by the very same GDR prompt/model that produces the refined pol100k dataset. Therefore the test set is not an external measure of whether the refined data preserved facts; it measures whether a model trained on GDR's cleaned text can answer questions that GDR itself generated from the original toxic text. The evaluation criterion is defined by the system under test, so the utility result is self-referential rather than an independent verification of fact preservation.

full rationale

The paper contains several genuinely independent and non-circular results: the large-scale code anonymization experiment (Section 5) uses 1.2M lines of real code with human-expert PII labels, and GDR is compared against DIRS on those labels; the toxicity reduction experiment (Section 6.1) uses the external Perspective API to score GDR's outputs; the diversity comparison (Section 7) is based on embedding distances; and the CompaniesQA utility test (Section 4.4) uses deterministically generated public/private labels. These provide external evidence that the GDR framework can work in practice. However, three evaluation components are self-referential: the PII benchmark in Section 4.1 is generated by the same model family as the GDR refiner, with prompts whose PII/non-PII cues align with the GDR prompt's own definition of PII; the SFT result in Section 4.3 trains and tests on the same procedural generator, making the reported improvement an in-sample fit; and pol5k-quiz in Section 6.2 is produced by the same GDR prompt that creates the refined dataset, so the utility check uses ground truth generated by the system under evaluation. These issues create partial circularity for headline-level claims about anonymization and detoxification utility, but they do not collapse the entire derivation because the code and toxicity results provide independent support. Self-citation is not load-bearing here; the Kirk et al. citation is motivational rather than central. Overall score 4 reflects partial circularity with substantial independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

GDR rests on empirical claims about LLM rewriting ability. The main burden is the synthetic benchmark and same-model evaluation, supplemented by hand-written prompts and automated toxicity scores. No mathematical derivation or formal privacy guarantee is provided.

axioms (5)
  • domain assumption A sufficiently capable pretrained LLM can identify and rewrite PII and toxic content in a single pass without specialized classifiers.
    This is the core premise of GDR, introduced in Section 3 and assumed throughout the experiments.
  • ad hoc to paper The synthetic PII benchmark generated via Gemini is representative of real-world PII leakage.
    Section 4.1 builds positives and negatives with Gemini and uses them to claim superiority over DIRS; no real-world PII corpus is used.
  • ad hoc to paper Fine-tuning on a GDR-refined dataset transfers world knowledge, as measured by pol5k-quiz extracted by the same GDR process.
    Section 6.2 constructs the test set from the same pipeline used to refine the data, which may overstate utility.
  • domain assumption Perspective API scores are a valid measure of toxicity for refined data.
    Section 6.1 uses Perspective API as the sole toxicity metric without human evaluation or alternative toxicity models.
  • domain assumption The expert annotations on the code dataset are ground truth for PII at line level.
    Section 5 treats human labels as ground truth, but the paper itself notes GDR sometimes flags PII missed by experts, so the labels are not fully objective.

pith-pipeline@v1.3.0-alltime-deepseek · 21386 in / 12299 out tokens · 126493 ms · 2026-08-04T20:15:28.788273+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Generative Data Refinement: Just Ask for Better Data." pith.science (2026). https://pith.science/paper/NK7FOLEI

@misc{pith2026250908653,
  author       = {Pith},
  title        = {Pith review of: Generative Data Refinement: Just Ask for Better Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NK7FOLEI}},
  note         = {Machine review of arXiv:2509.08653}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

For a fixed parameter size, the capabilities of large models are primarily determined by the quality and quantity of its training data. Consequently, training datasets now grow faster than the rate at which new data is indexed on the web, leading to projected data exhaustion over the next decade. Much more data exists as user-generated content that is not publicly indexed, but incorporating such data comes with considerable risks, such as leaking private information and other undesirable content. We introduce a framework, Generative Data Refinement (GDR), for using pretrained generative models to transform a dataset with undesirable content into a refined dataset that is more suitable for training. Our experiments show that GDR can outperform industry-grade solutions for dataset anonymization, as well as enable direct detoxification of highly unsafe datasets. Moreover, we show that by generating synthetic data that is conditioned on each example in the real dataset, GDR's refined outputs naturally match the diversity of web scale datasets, and thereby avoid the often challenging task of generating diverse synthetic data via model prompting. The simplicity and effectiveness of GDR make it a powerful tool for scaling up the total stock of training data for frontier models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection

    cs.AI 2026-06 unverdicted novelty 6.0

    Safety Reflection Pretraining adds regular safety reflections to pretraining data to integrate self-monitoring and reduce unsafe generalization from safe data in LLMs.

Reference graph

Works this paper leans on

54 extracted references · 28 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

  2. [2]

    Agarwal, A

    R. Agarwal, A. Singh, L. M. Zhang, B. Bohnet, S. Chan, A. Anand, Z. Abbas, A. Nova, J. D. Co-Reyes, E. Chu, et al. Many-shot in-context learning. arXiv preprint arXiv:2404.11018, 2024 a

  3. [3]

    Agarwal, N

    R. Agarwal, N. Vieillard, Y. Zhou, P. Stanczyk, S. R. Garea, M. Geist, and O. Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In The Twelfth International Conference on Learning Representations, 2024 b

  4. [4]

    C. Anil, E. Durmus, M. Sharma, J. Benton, S. Kundu, J. Batson, N. Rimsky, M. Tong, J. Mu, D. Ford, et al. Many-shot jailbreaking. Anthropic, April, 2024

  5. [5]

    Introducing claude

    Anthropic. Introducing claude. URL: https://www. anthropic. com/index/introducing-claude [accessed 2023-03-30], 2023

  6. [6]

    Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022

  7. [7]

    Bambauer, K

    J. Bambauer, K. Muralidhar, and R. Sarathy. Fool's gold: an illustrated critique of differential privacy. Vand. J. Ent. & Tech. L., 16: 0 701, 2013

  8. [8]

    Bhan, J.-N

    M. Bhan, J.-N. Vittaut, N. Achache, V. Legrand, N. Chesneau, A. Blangero, J. Murris, and M.-J. Lesot. Mitigating text toxicity with counterfactual generation. arXiv preprint arXiv:2405.09948, 2024

  9. [9]

    Boesel and R

    F. Boesel and R. Rombach. Improving image editing models with generative data refinement. In The Second Tiny Papers Track at ICLR 2024

  10. [10]

    Bradley, A

    H. Bradley, A. Dai, H. Teufel, J. Zhang, K. Oostermeijer, M. Bellagente, J. Clune, K. Stanley, G. Schott, and J. Lehman. Quality-diversity through ai feedback. arXiv preprint arXiv:2310.13032, 2023

  11. [11]

    Carlini, C

    N. Carlini, C. Liu, \'U . Erlingsson, J. Kos, and D. Song. The secret sharer: Evaluating and testing unintended memorization in neural networks. In 28th USENIX security symposium (USENIX security 19), pages 267--284, 2019

  12. [12]

    Carlini, F

    N. Carlini, F. Tramer, E. Wallace, M. Jagielski, A. Herbert-Voss, K. Lee, A. Roberts, T. Brown, D. Song, U. Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633--2650, 2021

  13. [13]

    D. Dale, A. Voronov, D. Dementieva, V. Logacheva, O. Kozlova, N. Semenov, and A. Panchenko. Text detoxification using large pre-trained neural models. arXiv preprint arXiv:2109.08914, 2021

  14. [14]

    Domingo-Ferrer, D

    J. Domingo-Ferrer, D. S \'a nchez, and A. Blanco-Justicia. The limits of differential privacy (and its misuse in data release and machine learning). Communications of the ACM, 64 0 (7): 0 33--35, 2021

  15. [15]

    Dubey, A

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  16. [16]

    C. Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1--12. Springer, 2006

  17. [17]

    L. Fan, K. Chen, D. Krishnan, D. Katabi, P. Isola, and Y. Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382--7392, 2024

  18. [18]

    L. Gao, S. Biderman, S. Black, L. Golding, T. Hoppe, C. Foster, J. Phang, H. He, A. Thite, N. Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020

  19. [19]

    L. Gao, J. Schulman, and J. Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835--10866. PMLR, 2023

  20. [20]

    Gemini Team , R. Anil, S. Borgeaud, Y. Wu, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

  21. [21]

    The mobile economy 2022, 2022

    GSMA. The mobile economy 2022, 2022

  22. [22]

    Gulcehre, T

    C. Gulcehre, T. L. Paine, S. Srinivasan, K. Konyushkova, L. Weerts, A. Sharma, A. Siddhant, A. Ahern, M. Wang, C. Gu, et al. Reinforced self-training (rest) for language modeling. arXiv preprint arXiv:2308.08998, 2023

  23. [23]

    Hoffmann, S

    J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. d. L. Casas, L. A. Hendricks, J. Welbl, A. Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022

  24. [24]

    Kaplan, S

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

  25. [25]

    J. H. Kirchner, Y. Chen, H. Edwards, J. Leike, N. McAleese, and Y. Burda. Prover-verifier games improve legibility of llm outputs. arXiv preprint arXiv:2407.13692, 2024

  26. [26]

    R. Kirk, I. Mediratta, C. Nalmpantis, J. Luketina, E. Hambro, E. Grefenstette, and R. Raileanu. Understanding the effects of rlhf on llm generalisation and diversity. arXiv preprint arXiv:2310.06452, 2023

  27. [27]

    Kumar, V

    A. Kumar, V. Zhuang, R. Agarwal, Y. Su, J. D. Co-Reyes, A. Singh, K. Baumli, S. Iqbal, C. Bishop, R. Roelofs, et al. Training language models to self-correct via reinforcement learning. arXiv preprint arXiv:2409.12917, 2024

  28. [28]

    Lange, Y

    R. Lange, Y. Tian, and Y. Tang. Large language models as evolution strategies. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, pages 579--582, 2024

  29. [29]

    Laugier, J

    L. Laugier, J. Pavlopoulos, J. Sorensen, and L. Dixon. Civil rephrases of toxic texts with self-supervised transformers. arXiv preprint arXiv:2102.05456, 2021

  30. [30]

    J. Lee, Z. Dai, X. Ren, B. Chen, D. Cer, J. R. Cole, K. Hui, M. Boratko, R. Kapadia, W. Ding, et al. Gecko: Versatile text embeddings distilled from large language models. arXiv preprint arXiv:2403.20327, 2024

  31. [31]

    A. Lees, V. Q. Tran, Y. Tay, J. Sorensen, J. Gupta, D. Metzler, and L. Vasserman. A new generation of Perspective API : Efficient multilingual character-level transformers. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 3197--3207, 2022

  32. [32]

    Lehman, J

    J. Lehman, J. Gordon, S. Jain, K. Ndousse, C. Yeh, and K. O. Stanley. Evolution through large models. In Handbook of Evolutionary Machine Learning, pages 331--366. Springer, 2023

  33. [33]

    Lewis, D

    M. Lewis, D. Yarats, Y. N. Dauphin, D. Parikh, and D. Batra. Deal or no deal? end-to-end learning for negotiation dialogues. arXiv preprint arXiv:1706.05125, 2017

  34. [34]

    L. Li, L. Fan, S. Atreja, and L. Hemphill. “hot” chatgpt: The promise of chatgpt in detecting and discriminating hateful, offensive, and toxic comments on social media. ACM Transactions on the Web, 18 0 (2): 0 1--36, 2024

  35. [35]

    C.-Y. Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81, 2004

  36. [36]

    L. Long, R. Wang, R. Xiao, J. Zhao, X. Ding, G. Chen, and H. Wang. On llms-driven synthetic data generation, curation, and evaluation: A survey. Findings of the Association for Computational Linguistics, 2024

  37. [37]

    Lukas, A

    N. Lukas, A. Salem, R. Sim, S. Tople, L. Wutschitz, and S. Zanella-B \'e guelin. Analyzing leakage of personally identifiable information in language models. In 2023 IEEE Symposium on Security and Privacy (SP), pages 346--363. IEEE, 2023

  38. [38]

    Lupidi, C

    A. Lupidi, C. Gemmell, N. Cancedda, J. Dwivedi-Yu, J. Weston, J. Foerster, R. Raileanu, and M. Lomeli. Source2synth: Synthetic data generation and curation grounded in real data sources. arXiv preprint arXiv:2409.08239, 2024

  39. [39]

    Maini, S

    P. Maini, S. Seto, H. Bai, D. Grangier, Y. Zhang, and N. Jaitly. Rephrasing the web: A recipe for compute and data-efficient language modeling. arXiv preprint arXiv:2401.16380, 2024

  40. [40]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018

  41. [41]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

  42. [42]

    Papasavva, S

    A. Papasavva, S. Zannettou, E. De Cristofaro, G. Stringhini, and J. Blackburn. Raiders of the lost kek: 3.5 years of augmented 4chan posts from the politically incorrect board. In Proceedings of the international AAAI conference on web and social media, volume 14, pages 885--894, 2020

  43. [43]

    Pavlopoulos, J

    J. Pavlopoulos, J. Sorensen, L. Dixon, N. Thain, and I. Androutsopoulos. Toxicity detection: Does context really matter? arXiv preprint arXiv:2006.00998, 2020

  44. [44]

    Email statistics report, 2021-2025

    Radicati Group. Email statistics report, 2021-2025. Executive summary, The Radicati Group, Inc., 2020

  45. [45]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024

  46. [46]

    M. Reid, N. Savinov, D. Teplyashin, D. Lepikhin, T. Lillicrap, J.-b. Alayrac, R. Soricut, A. Lazaridou, O. Firat, J. Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024

  47. [47]

    Samvelyan, S

    M. Samvelyan, S. C. Raparthy, A. Lupu, E. Hambro, A. H. Markosyan, M. Bhatt, Y. Mao, M. Jiang, J. Parker-Holder, J. Foerster, et al. Rainbow teaming: Open-ended generation of diverse adversarial prompts. arXiv preprint arXiv:2402.16822, 2024

  48. [48]

    Schwartz

    O. Schwartz. In 2016, microsoft’s racist chatbot revealed the dangers of online conversation. IEEE spectrum, 11: 0 2019, 2019

  49. [49]

    S. Song, K. Chaudhuri, and A. D. Sarwate. Stochastic gradient descent with differentially private updates. In 2013 IEEE global conference on signal and information processing, pages 245--248. IEEE, 2013

  50. [50]

    Staab, M

    R. Staab, M. Vero, M. Balunovi \'c , and M. Vechev. Beyond memorization: Violating privacy via inference with large language models. arXiv preprint arXiv:2310.07298, 2023

  51. [51]

    Thomas and A

    M. Thomas and A. T. Joy. Elements of information theory. Wiley-Interscience, 2006

  52. [52]

    Villalobos, J

    P. Villalobos, J. Sevilla, L. Heim, T. Besiroglu, M. Hobbhahn, and A. Ho. Will we run out of data? an analysis of the limits of scaling datasets in machine learning. arXiv preprint arXiv:2211.04325, 2022

  53. [53]

    H. Xin, D. Guo, Z. Shao, Z. Ren, Q. Zhu, B. Liu, C. Ruan, W. Li, and X. Liang. Deepseek-prover: Advancing theorem proving in llms through large-scale synthetic data. arXiv preprint arXiv:2405.14333, 2024

  54. [54]

    Z. Yang, N. Band, S. Li, E. Cand \`e s, and T. Hashimoto. Synthetic continued pretraining. arXiv preprint arXiv:2409.07431, 2024