- Rating
- 4(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies the contribution of fiction to language-model pretraining through controlled experiments with 1B-parameter models trained on 185B tokens. It compares a web-only baseline with mixtures containing contemporary Books3 fiction, public-domain Institutional Books fiction, and synthetic fiction generated from Books3. Contemporary fiction improves automated creative-writing performance and substantially outperforms the selected public-domain corpus, while offering less consistent benefits for other capabilities. Synthetic fiction preserves or improves several writing scores while reducing the effectiveness of tested membership-inference and extraction attacks against the original books. The contribution is empirical evidence connecting fiction-source selection, downstream utility, and source detectability.
Critical strengths
-
The controlled experiments establish useful differences between fiction sources. Unlike comparisons of unrelated pretrained models, the study holds major training choices fixed while changing the fiction component. Books3 achieves EQ-Bench win rates of 78.6 ± 3.9% against NoBooks and 91.1 ± 4.5% against Institutional Books.
-
The paper jointly examines utility and detectability after synthetic transformation. Synthetic achieves a 79.7 ± 6.8% EQ-Bench win rate against Books3, while the tested attacks become less effective at detecting or reproducing the original text.
Critical weaknesses
-
The strongest creative-writing conclusions lack adequate human validation. Table 9 reports a 57% human win rate for Books3 against NoBooks over 50 examples, with one annotation per example and no reported uncertainty. Human–judge agreement is 46% for both LitBench comparisons. Agreement in the direction of aggregate preferences provides some corroboration, but does not establish reliable measurement of the quality differences.
-
More importantly, the human study does not evaluate Synthetic, despite its particularly large automated gains. This leaves uncertainty about whether its advantage translates into reader preference. A small blinded comparison using existing Synthetic and Books3 outputs would directly validate a central result without expanding the paper’s scope.
-
The provenance conclusions extend beyond what the reported experiment establishes. Figure 3 demonstrates reduced detectability under four membership-inference methods and one extraction procedure. It does not establish that provenance auditing generally becomes ineffective. The original books are sources of synthetic-data generation rather than literal training documents, so this distinction is important when interpreting attack failure.
-
§4.2 discusses results across context lengths without specifying the lengths or their aggregation in Figure 3a. Its expectation that the extraction gap will increase at larger scales is also untested.
Questions for authors
-
Targeted validation: Can you provide a blinded human comparison of Synthetic versus Books3 using a modest sample of existing generations, with win/loss/tie counts and uncertainty estimates? This would strengthen the central utility claim.
-
Auditing clarification and control: What context lengths and passage-to-document aggregation procedures produced Figure 3a, and how were the member/non-member books separated from synthetic-generation inputs? Reporting the same attacks on the existing NoBooks model would help assess whether the candidate sets are distinguishable without the added Books3 exposure.
-
Training-data clarification: Appendix G.2 reports approximately 450,000 generated chapters totaling 6B tokens, implying about 13,333 tokens per chapter, whereas §4.1 targets 1,000–2,000 words. Could you reconcile these figures and clarify whether the final 5B synthetic training tokens contain repeated generations? This matters for interpreting the matched-budget comparison.
4 comments
Continuation of Author Response to Reviewer RhCp
[9] “Scalable Extraction of Training Data from (Production) Language Models” (Nasr et al., ICLR 2025).
[10] “Extracting memorized pieces of (copyrighted) books from open-weight language models” (Cooper et al., CoLM 2026).
[11] “Extracting books from production language models” (Cooper et al., 2026).
Auditing clarification and control: What context lengths and passage-to-document aggregation procedures produced Figure 3a, and how were the member/non-member books separated from synthetic-generation inputs? Reporting the same attacks on the existing NoBooks model would help assess whether the candidate sets are distinguishable without the added Books3 exposure.
We’ll try to answer each of the reviewer’s subquestions here one at a time:
how were the member/non-member books separated from synthetic-generation inputs
The reviewer's questions helped us identify a possible data leak in the nonmember set used specifically for evaluating the synthetic-data model where some candidate nonmember books may have served as conditioning inputs during synthetic-story generation. This possibility does not affect the Books3 or NoBooks evaluations. To eliminate it, we replaced the nonmember set with external BookMIA nonmembers and reran all attacks (across all models and seeds) at context lengths 128, 256, and 512. We omit 1,024 because BookMIA does not contain enough continuous nonmember passages of that length for a matched evaluation. The revised results support the same conclusions as before. We also add this discussion in Appendix K.1.
What context lengths and passage-to-document aggregation procedures produced Figure 3a
Figure 3a reports passage-level AUROC at a context length of 512 tokens. We have also now added additional sequence length details in Appendix K in the revised manuscript and thank the reviewer for bringing this to our attention. We do not perform any passage-to-document aggregation. All member passages are samples from different documents.
Reporting the same attacks on the existing NoBooks model would help assess whether the candidate sets are distinguishable without the added Books3 exposure.
We run the membership inference and extraction attacks on the existing 185B-token NoBooks model using three training seeds. Mean AUROC values were:
| Attack | 128 | 256 | 512 |
|---|---|---|---|
| Loss | 0.507 | 0.500 | 0.505 |
| Min-K | 0.489 | 0.480 | 0.486 |
| Min-K++ | 0.528 | 0.505 | 0.497 |
| Zlib | 0.407 | 0.413 | 0.431 |
All four attacks on the NoBooks model perform close to chance across context lengths (Appendix K). At 512 tokens, the Books3-exposed model achieves AUROCs of 0.703 (Loss), 0.685 (Min-K), 0.655 (Min-K++), and 0.705 (Zlib), whereas the synthetic-exposed model achieves 0.515 (Loss), 0.510 (Min-K), 0.552 (Min-K++), and 0.455 (Zlib). Thus, the strong attack performance is specific to the Books3-exposed model; the synthetic-exposed and NoBooks models remain close to chance.
In addition, we run our training data extraction attack (measured via BLEU) on the NoBooks model as baseline. We find that Books3 data is no more extractable from the synthetic model than from the No Books model, which was never trained on Books3. We have updated the paper to reflect these results (Appendix K.2).
| Model | Mean BLEU | Rate BLEU ≥ 0.75 |
|---|---|---|
| Books3 | 0.0433 ± 0.0001 | 2.27e-04 ± 1e-05 |
| No Books | 0.0388 ± 0.0000 | 7.67e-05 ± 2e-05 |
| Synthetic | 0.0388 ± 0.0000 | 6.33e-05 ± 1e-05 |
§4.2 discusses results across context lengths without specifying the lengths or their aggregation in Figure 3a. Its expectation that the extraction gap will increase at larger scales is also untested.
It is well-established in memorization literature that memorization rate rises with model scale [1], hence our expectation that the extraction gap will get larger. On the sequence length, Fig 3a now uses 512 token sequences however we also experiment with lengths of 128 and 256. We have added them under Appendix K.1.
[1] “Quantifying Memorization Across Neural Language Models” (Carlini et al., ICLR 2023).
Appendix G.2 reports approximately 450,000 generated chapters totaling 6B tokens, implying about 13,333 tokens per chapter, whereas §4.1 targets 1,000–2,000 words.
Appendix G.2 contained a typo which is now fixed in the updated manuscript. Thank you for bringing it to our attention. The correct number of generated chapters is ~4500k. We first generated ~6B tokens of data from such synthetic chapters. We then remove any chapters containing at least a 50-word overlap with Books3 data (L 357-358). This procedure removes ~24M tokens. From the remaining ~5.98B tokens, we randomly sample 5B tokens for pretraining and discard the rest.
I thank the authors for their comprehensive revision. I particularly appreciate the additional human evaluation and the NoBooks control results. These additions address several of my concerns, and I have updated my score accordingly.
Response to Reviewer Comment
Thank you for taking the time to engage with our work and give us feedback!
Author Response to Reviewer RhCp
We are grateful that the reviewer appreciated how our controlled experimental setup could help establish differences in quality between sources of fiction. We are also glad the reviewer valued how our work brings light to both the utility of synthetic data and detectability of the original source material for that synthetic data. Below we address all the points in your review, and remain available to address any remaining questions you may have before the end of the response period.
We note that the LitBench reward model is an established human-validated measure from prior work [1], rather than a new LLM-as-a-judge method we have constructed for this work. In [1], the reward model was validated on a held-out set of 2,480 human-labeled story comparisons and reached 78% agreement with human preferences. We use this reward model for our experiments, adding our own human annotations in the appendix only for completeness.
We add a second human annotator so that every example now has two independent annotations. We report inter-annotator agreement and 95% confidence intervals. On LitBench, the two annotators agree on 64% of examples. The human win rate for Books3 vs No Books is 57% for the first annotator and 61% for the second. The preference for Books3 is consistent across both annotators. The human and judge agreement is 46% and 47% for two annotators. We have updated our paper to reflect these results (Appendix J and Table 9).
[1] LitBench: A Benchmark and Dataset for Reliable Evaluation of Creative Writing (Fein et al., EACL 2026)
We use the LitBench reward model for these experiments as well, which is an established human-validated measure from prior work [1], rather than a new LLM-as-a-judge method we propose. In [1], the reward model was validated on a held-out set of 2,480 human-labeled story comparisons and reached 78% agreement with human preferences. We use this reward model for our experiments, adding our own human annotations in the appendix only for completeness. We additionally conduct a blind human comparison of Synthetic vs Book3 generations for 50 examples on LitBench, with two independent annotations per example. Across both annotators, Synthetic wins 35, Books3 wins 42, and 23 are ties. This gives the synthetic model a win rate of 46.5% (95% CI 36-58%). These results suggest the synthetic model is statistically comparable to Books3 in human reader preference. The revised manuscript reflects these results (Appendix J and Table 9).
[1] LitBench: A Benchmark and Dataset for Reliable Evaluation of Creative Writing (Fein et al., EACL 2026).
We show that it is possible to generate high-quality synthetic data by conditioning on Books3 that can give similar performance improvements to training on the original source material. We demonstrate that some of the most prominent membership inference and data extraction attacks available are not successful at detecting use of the original copyrighted material as a source for this synthetic data in the resulting model. We emphasize that the membership inference and data extraction attacks that we use in the paper represent some of the most widely-used attacks in the field. For membership inference: the loss-based [1], Min-K% [2], Min-K%++ [3], and zLib [4] attacks have been widely used to establish provenance for pre-training data [5, 6, 7]. For data extraction: the method that we use in the paper has been extensively used by prior works for extracting pre-training data [4, 8], including from production models [9] and for copyrighted books [10, 11].
[1] “Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting” (Yeom et al., IEEE CSF 2018).
[2] “Detecting Pretraining Data From Large Language Models” (Shi et al., ICLR 2024).
[3] “Improved baseline for detecting pre-training data from large language models” (Zhang et al., ICLR 2025).
[4] “Extracting Training Data From Large Language Models” (Carlini et al., USENIX 2021).
[5] “RECALL: Membership Inference via Relative Conditional Log-Likelihoods” (Xie et al., EMNLP 2024).
[6] “CON-RECALL: Detecting Pre-training Data in LLMs via Contrastive Decoding” (Wang et al., COLING 2025).
[7] “Hubble: A Model Suite to Advance the Study of LLM Memorization” (Wei et al., ICLR 2026).
[8] “Quantifying Memorization Across Neural Language Models” (Carlini et al., ICLR 2023)