Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Libra: Large Chinese-based Safeguard for AI Content

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An open-source Chinese content safeguard reaches 86.79% accuracy, beating open baselines and nearing closed commercial moderators.

desk verdict A useful Chinese safety model and benchmark, but the headline numbers rest on a benchmark that shares its data source with training; the external translated split saves the core claim. read the letter →

arxiv 2507.21929 v1 pith:EB5WNWEA submitted 2025-07-29 cs.AI

Ziyang Chen , Huimu Yu , Xing Wu , Dongqin Liu , Songlin Hu This is my paper
classification cs.AI
keywords Chinesecontentmoderationguardmodeltwo-stagecurriculumtrainingsyntheticdataLLMsafetybenchmarkevaluationred-teaming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a Chinese-language content safeguard can be built without heavy manual annotation: first pretrain a guard model on about 240,000 synthetic query-response pairs, then fine-tune on roughly 18,000 real, harder examples. The result, Libra-Guard, reaches 86.79% average accuracy on the accompanying Libra-Test benchmark, outperforming open-source alternatives such as Qwen2.5-14B-Instruct and ShieldLM-Qwen-14B-Chat, and approaching closed-source models like GPT-4o and Claude-3.5-Sonnet. Libra-Test is presented as the first benchmark built specifically for evaluating Chinese safeguards, with over 5,700 samples across seven harm categories. If these claims hold, open-source Chinese LLM deployments can get commercial-level moderation at lower cost and with full data control.

What carries the argument

The two-stage curriculum training pipeline: Stage 1, guard pretraining, synthesizes ~240k harmful queries (extending AART with Chinese-specific harmful events), generates responses with several base and instruct models, retains only samples where Qwen-2.5-32B-Instruct and Gemma-2-27B-it agree on the safety label, and balances safe and unsafe responses. Stage 2, guard finetuning, extracts real harmful queries from Safety-Prompts, generates responses with the same generator pool, uses weak models to flag inconsistent labels, and relabels those hard samples with GPT-4o before full-parameter fine-tuning for one epoch. A central design detail is the Rear Critic: during training the model produces an analysis plus the label, with the critic placed after the label, which improves average accuracy from 81.80% (No Critic) to 86.48% in ablations. The evaluation benchmark takes three data streams (real, synthetic, and translated from BeaverTails), keeps samples where two open models disagree, and labels them by three human annotators plus a safety expert, yielding a balanced mix of 1,864 safe and 3,856 unsafe examples.

What would settle it

Evaluate Libra-Guard on a newly collected, independently annotated set of Chinese harmful queries and responses that are disjoint from Safety-Prompts and other public red-team datasets. If its accuracy advantage over Qwen2.5-14B-Instruct and ShieldLM-Qwen-14B-Chat shrinks to near the base-model gaps, the claim of general superiority would be undercut. A simpler check: measure n-gram or embedding overlap between the finetuning queries and Libra-Test's real and synthetic samples; substantial overlap would indicate the benchmark is partially a memorization test.

Watch

Extended reading notes

Core claim

Libra-Guard, when built on Qwen2.5-14B-Instruct, achieves 86.79% accuracy on Libra-Test, with F1-safe of 80.64% and F1-unsafe of 89.83%. The same training recipe transfers to smaller bases: 0.5B, 1.5B, 3B, 7B, and even a 0.1B discriminative model (Libra-Tiny) all exceed their starting instruct models and most baseline guards. The paper attributes this to the two-stage curriculum: synthetic adversarial pretraining builds a broad foundation, and fine-tuning on hard, real-world examples sharpens it. It also reports that placing the critic explanation after the label (Rear Critic) outperforms No Critic and Front Critic, that explicit safety rules in the prompt are unnecessary after this training, and that pretraining improves sample efficiency—pretrained models outperform from-scratch fine-tuning at every data size. The central claim is that this pipeline yields a Chinese safeguard approaching closed-source proprietary moderators while staying fully open.

Load-bearing premise

The load-bearing premise is that Libra-Test measures general Chinese safety ability rather than overlap with Libra-Guard's own training distribution—the real test samples come from Safety-Prompts, which also supplies the finetuning queries, and the synthetic test construction is referenced to a nonexistent section.

Editorial extensions

If this is right

  • Open-source Chinese LLM deployments can integrate a moderating gate that approaches proprietary APIs without sending user content off-premise.
  • The synthetic-pretraining-plus-real-finetuning recipe may cut annotation cost for safety systems in other non-English or domain-specific contexts.
  • Small models get a large boost: the 0.5B version reaches 81.46% accuracy, suggesting edge-deployable content moderation is feasible.
  • A benchmark dedicated to Chinese safeguards (Libra-Test) will let future systems be compared on the same seven harm categories.
  • The finding that explicit safety rules add little once critic-based training is in place aligns with reasoning-based alignment approaches and may simplify deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the accuracy gap over base models holds on fresh data, much of Chinese moderation ability may be learnable from synthetic English-style red-teaming plus a small real Chinese set, lowering the data barrier for other languages.
  • A direct test of the paper's generality claim would be to apply the same pipeline to another under-resourced language and measure whether the baseline gap persists.
  • The rear-critic advantage hints that classifier training benefits from explanation-after-decision; applying this ordering to other safety and classification tasks could be tested independently.
  • Because the real-data portion of Libra-Test shares its source with the finetuning queries, the benchmark's independence from training remains an open question that only an external, disjoint evaluation can settle.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Libra-Guard, a Chinese-language content safeguard built with a two-stage curriculum pipeline: large-scale synthetic-data pretraining followed by fine-tuning on real-world queries, and Libra-Test, a benchmark of 5,720 Chinese safety samples spanning real, synthetic, and translated sources. On Libra-Test, Libra-Guard-Qwen2.5-14B-Instruct reports 86.79% accuracy, surpassing Qwen2.5-14B-Instruct (74.33%) and ShieldLM-Qwen-14B-Chat (65.69%) and approaching GPT-4o (91.05%) and Claude-3.5-Sonnet (88.82%). The paper also reports ablations on scaling, critic placement, safety rules, curriculum order, and annotation-model agreement, plus an appendix result for a 0.1B discriminative model.

Significance. If the reported evaluation is trustworthy, the paper makes a useful contribution: an open-source, Chinese-specific guard model with a scalable data generation pipeline, and a new benchmark that can support future work on Chinese content safety. The paper's strengths include the public release of model and benchmark, the breadth of ablation studies (Sections 5.1-5.5), the inclusion of multiple base models, and the use of a seven-category safety taxonomy with human annotation. However, the significance of the headline claim depends heavily on Libra-Test being an independent measure of general Chinese safety ability, and that independence is not currently established: the real-data split shares its source with the fine-tuning queries, and the synthetic-split construction is referenced to a nonexistent section. The benchmark and the guarded model are still likely to be useful resources, but the paper's central comparison should be verified against an external or demonstrably disjoint evaluation set.

major comments (4)
  1. [Section 3.1 vs. Section 2.2] The central claim that Libra-Guard outperforms open-source baselines by 12-21 percentage points (Table 2) depends on Libra-Test being an independent measure of Chinese safeguard quality. However, Section 3.1 states that the real-data portion of Libra-Test is drawn from Safety-Prompts [22], and Section 2.2 states that fine-tuning queries are randomly extracted from Safety-Prompts. The paper only asserts 'no overlap' without describing any overlap-removal procedure or providing exact disjointness evidence. Sample-level non-overlap between two random draws does not rule out distributional leakage: a model trained on one draw from Safety-Prompts can be expected to score higher on a held-out draw from the same source than baselines that were never trained on that source. The authors should provide explicit instance-level disjointness verification and, ideally, evaluate on a Chinese safety test set that is external to Safety-Prompts or use a cross-validation-style protocol where the test split is excluded from any training-data selection.
  2. [Section 3.1 (reference to nonexistent Section 2.3)] The synthetic-data portion of Libra-Test is described as 'detailed in Section 2.3,' but Section 2 contains only Sections 2.1 and 2.2. If the intended reference is Section 2.1, then the synthetic test items are generated by the same Qwen-based synthesis pipeline used to create the pretraining instances, placing the test items inside the training distribution. If the intended reference is something else, the construction details are missing and the synthetic split cannot be reproduced or audited. This is not a presentation nitpick: the synthetic split accounts for 1,467 of the 5,720 test samples, so an unverifiable or training-overlapping construction materially weakens the headline accuracy and the reported baseline gaps. The authors should either add the missing section or, if the test items are built differently from the training data, specify exactly how they differ and provide the generation prompts and deduplication procedure.
  3. [Section 4.2, Table 2] No confidence intervals, standard deviations, or significance tests are reported for any of the accuracy or F1 numbers. The headline gaps (86.79% vs. 74.33% vs. 65.69%) are large, so the qualitative conclusion may survive, but the paper does not quantify evaluation uncertainty. Because the test set is 5,720 samples and the F1-Safe/F1-Unsafe values vary widely across baselines (e.g., F1-Safe of 6.47% for Qwen2.5-0.5B-Instruct), the authors should report bootstrap confidence intervals or a significance test, and should state the evaluation protocol used for all models (e.g., same prompt template, decoding hyperparameters, number of samples per query). This is needed to determine whether the observed differences are reliable or partly reflect prompt-format and decoding sensitivity.
  4. [Section 2.2 and 3.2 (annotation pipeline overlap)] The test set's 'hard' real and synthetic samples are selected by keeping responses where Qwen-2.5-32B-Instruct and Gemma-2-27B-it give inconsistent labels, and those same two models are used in the training-data annotation pipeline (Section 2.1 uses their agreement; Section 2.2 uses their disagreement to select hard fine-tuning samples). This means the benchmark's difficulty criterion is defined by the same annotation models used to construct training data, so the benchmark may implicitly favor a guard model trained to match those annotators' error patterns. The authors should clarify whether the test-set inclusion criterion interacts with the training-set selection criteria, and should consider reporting performance on samples that were not filtered through these two annotators.
minor comments (5)
  1. [Section 3.1] The third bullet, 'Translated Data', is missing the bullet formatting and appears as part of the previous sentence; this makes the three data sources harder to read.
  2. [Section 5.1] The phrase 'at the most enormous scale' should be replaced with 'at the largest scale', and 'outperform that textitwithout' contains a malformed LaTeX fragment that should be fixed to 'outperform those without'.
  3. [Section 6] The related-work text contains the apparent typo 'AeigsSafetTest'; it should read 'AegisSafetyTest' or similar.
  4. [Abstract and Section 1] The claim that Libra-Test is 'the first benchmark specifically designed to evaluate' Chinese safeguard systems should be scoped, since ShieldLM [30] already evaluates Chinese safety and the abstract's 'first benchmark' phrasing is stronger than what Section 3 supports.
  5. [Appendix C] The Libra-Tiny result (77.63% accuracy on 0.1B parameters) is reported only in an appendix and is not mentioned in the main results or contribution list; if it is intended as a contribution, it should be linked to the main narrative, and if it is only a side observation, the appendix should say so.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the safety-tuning and evaluation use a held-out split from a common source, not the same examples, and no claim reduces to its inputs by construction.

full rationale

The only place a reader might suspect circularity is the relationship between Libra-Guard's training data and Libra-Test's real-data split. Section 2.2 states that finetuning queries are 'randomly extracted from Safety-Prompts' while 'ensuring no overlap with the real data used in the Libra-Test,' and Section 3.1 says the real-data portion is also 'from the Safety-Prompts dataset.' This is a held-out split from a common dataset rather than training on the test examples, which is standard evaluation practice and does not make the reported 86.79% accuracy equivalent to the training input. The synthetic-data split is harder to audit: Section 3.1 refers to a nonexistent Section 2.3 for its construction, so the paper does not currently show that synthetic test queries are independent of the 240k pretraining synthetic queries. That is a verifiability and missing-support flaw in the benchmark-independence claim, but it is not itself a demonstrated circular reduction; without a quoted construction tying the test samples back to training inputs, the specific reduction required for a circularity finding is absent. Self-citations, such as [28] (CodePMP) in the two-stage-training inspiration and [12] (SPC) in Appendix C, are incidental and not load-bearing. The translated BeaverTails split is external and provides independent evidence. Therefore no circularity score above 0 is warranted.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the completeness of the seven safety rules, the reliability of model-based and human annotation, and the similarity between synthetic training and test generation. None of these are independently established in the paper, which is normal for an empirical ML submission but should be checked.

free parameters (3)
  • Learning rate = not reported
    Section 4.1 specifies Adam and a linear decay schedule but omits the learning rate, leaving a key training hyperparameter undefined.
  • Annotation agreement threshold = two-model consistency
    Section 2.1 keeps only synthetic samples where Qwen-2.5-32B-Instruct and Gemma-2-27B-it assign the same label, a design choice that shapes the training distribution.
  • Safe/unsafe sampling ratio = 1:1 per query
    Section 2.1 samples one safe and one unsafe response per query, setting the class balance for the 240k pretraining set.
assumptions (4)
  • domain assumption The seven safety rules define a complete and correct taxonomy of harmful Chinese content.
    Appendix A; the benchmark and model both rely on these rules, so errors in the taxonomy propagate to labels and scores.
  • domain assumption Majority vote of three annotators with expert confirmation yields ground truth.
    Section 3.3; no inter-annotator agreement is reported, so label reliability is assumed.
  • domain assumption Open-source models (Qwen-2.5-32B-Instruct, Gemma-2-27B-it) can reliably annotate safety responses.
    Section 2.1; used to label 240k pretraining samples without human verification.
  • domain assumption Synthetic adversarial queries generated by Qwen-2.5-32B-Instruct cover the space of Chinese harmful content.
    Section 2.1; the diversity of training and test data depends on this generator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Libra: Large Chinese-based Safeguard for AI Content." pith.science (2026). https://pith.science/paper/EB5WNWEA

@misc{pith2026250721929,
  author       = {Pith},
  title        = {Pith review of: Libra: Large Chinese-based Safeguard for AI Content},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EB5WNWEA}},
  note         = {Machine review of arXiv:2507.21929}
}
read the original abstract

Large language models (LLMs) excel in text understanding and generation but raise significant safety and ethical concerns in high-stakes applications. To mitigate these risks, we present Libra-Guard, a cutting-edge safeguard system designed to enhance the safety of Chinese-based LLMs. Leveraging a two-stage curriculum training pipeline, Libra-Guard enhances data efficiency by employing guard pretraining on synthetic samples, followed by fine-tuning on high-quality, real-world data, thereby significantly reducing reliance on manual annotations. To enable rigorous safety evaluations, we also introduce Libra-Test, the first benchmark specifically designed to evaluate the effectiveness of safeguard systems for Chinese content. It covers seven critical harm scenarios and includes over 5,700 samples annotated by domain experts. Experiments show that Libra-Guard achieves 86.79% accuracy, outperforming Qwen2.5-14B-Instruct (74.33%) and ShieldLM-Qwen-14B-Chat (65.69%), and nearing closed-source models like Claude-3.5-Sonnet and GPT-4o. These contributions establish a robust framework for advancing the safety governance of Chinese LLMs and represent a tentative step toward developing safer, more reliable Chinese AI systems.

Figures

Figures reproduced from arXiv: 2507.21929 by the authors.

Figure 1
Figure 1. Overview of the two-stage curriculum training for Libra-Guard. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. It targets three key aspects: diversity, difficulty, and consistency. Table 1 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 2
Figure 2. Overview of the construction process of the Libra-Test. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Scaling effects of pretraining (left) and finetuning (right): pretraining [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 12 canonical work pages

  1. [22]

    arXiv preprint arXiv:2304.10436 (2023)

    Sun, H., Zhang, Z., Deng, J., Cheng, J., Huang, M.: Safety assessment of chinese large language models. arXiv preprint arXiv:2304.10436 (2023)

    arXiv 2023

  2. [1]

    arXiv preprint arXiv:2303.08774 (2023)

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

    arXiv 2023

  3. [2]

    Claude-3 Model Card1 (2024)

    Anthropic, A.: The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card1 (2024)

    work page 2024

  4. [3]

    arXiv preprint arXiv:2112.00861 (2021)

    Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 (2021)

    arXiv 2021

  5. [4]

    In: Proceedings of the 26th annual international conference on machine learning

    Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41–48 (2009)

    2009

  6. [5]

    In: Proceed- ings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region

    Deng, Y., Lei, W., Huang, M., Chua, T.S.: Rethinking conversational agents in the era of llms: Proactivity, non-collaborativity, and beyond. In: Proceed- ings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region. pp. 298– 301 (2023)

    work page 2023

  7. [6]

    arXiv preprint arXiv:2405.19846 (2024)

    Gao, C., Wu, X., Fu, Q., Hu, S.: Quest: Query-centric data synthesis ap- proach for long-context scaling of large language model. arXiv preprint arXiv:2405.19846 (2024)

    arXiv 2024

  8. [7]

    Gao, C., Wu, X., Lin, Z., Zhang, D., Hu, S.: Longmagpie: A self-synthesis method for generating large-scale long-context instructions (2025), https: //arxiv.org/abs/2505.17134

    arXiv 2025

Show all 37 references
  1. [8]

    Gao, C., Wu, X., Lin, Z., Zhang, D., Hu, S.: Nextlong: Toward effective long-context training without long documents (2025), https://arxiv.org/ abs/2501.12766

    2025 arXiv

  2. [9]

    arXiv preprint arXiv:2404.05993 (2024)

    Ghosh, S., Varshney, P., Galinkin, E., Parisien, C.: Aegis: Online adaptive ai content safety moderation with ensemble of llm experts. arXiv preprint arXiv:2404.05993 (2024)

    2024 arXiv

  3. [10]

    arXiv preprint arXiv:2412.16339 (2024)

    Guan, M.Y., Joglekar, M., Wallace, E., Jain, S., Barak, B., Heylar, A., Dias, R., Vallone, A., Ren, H., Wei, J., et al.: Deliberative alignment: Reasoning enables safer language models. arXiv preprint arXiv:2412.16339 (2024)

    2024 arXiv

  4. [11]

    arXiv preprint arXiv:2406.18495 (2024)

    Han, S., Rao, K., Ettinger, A., Jiang, L., Lin, B.Y., Lambert, N., Choi, Y., Dziri, N.: Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms. arXiv preprint arXiv:2406.18495 (2024)

    2024 arXiv

  5. [12]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Hu, D., Wei, L., Liu, Y., Zhou, W., Hu, S.: Structured probabilistic coding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 12491–12501 (2024)

    2024

  6. [13]

    arXiv preprint arXiv:2410.21276 (2024)

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    2024 arXiv

  7. [14]

    Inan, H., Upasani, K., Chi, J., Rungta, R., Iyer, K., Mao, Y., Tontchev, M., Hu, Q., Fuller, B., Testuggine, D., et al.: Llama guard: Llm- 12 Z. Chen, H. Yu, X. Wu et al. based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674 (2023)

    2023 arXiv

  8. [15]

    Advances in Neural Information Processing Systems 36, 24678–24704 (2023)

    Ji, J., Liu, M., Dai, J., Pan, X., Zhang, C., Bian, C., Chen, B., Sun, R., Wang, Y., Yang, Y.: Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems 36, 24678–24704 (2023)

    2023

  9. [16]

    Li, Z.Z., Zhang, D., Zhang, M.L., Zhang, J., Liu, Z., Yao, Y., Xu, H., Zheng, J., Wang, P.J., Chen, X., Zhang, Y., Yin, F., Dong, J., Li, Z., Bi, B.L., Mei, L.R., Fang, J., Liang, X., Guo, Z., Song, L., Liu, C.L.: From system 1 to system 2: A survey of reasoning large language...

    2025 arXiv

  10. [17]

    arXiv preprint arXiv:2406.16694 (2024)

    Liang, X., Hu, X., Zuo, S., Gong, Y., Lou, Q., Liu, Y., Huang, S.L., Jiao, J.: Task oriented in-domain data augmentation. arXiv preprint arXiv:2406.16694 (2024)

    2024 arXiv

  11. [18]

    arXiv preprint arXiv:2506.08989 (2025)

    Liang, X., Li, Z.Z., Gong, Y., Wang, Y., Zhang, H., Shen, Y., Wu, Y.N., Chen, W.: Sws: Self-aware weakness-driven problem synthesis in reinforce- ment learning for llm reasoning. arXiv preprint arXiv:2506.08989 (2025)

    2025 arXiv

  12. [19]

    arXiv preprint arXiv:2401.02777 (2024)

    Liu, N., Chen, L., Tian, X., Zou, W., Chen, K., Cui, M.: From llm to conversational agent: A memory enhanced architecture with fine-tuning of large language models. arXiv preprint arXiv:2401.02777 (2024)

    2024 arXiv

  13. [20]

    arXiv preprint arXiv:2402.04249 (2024)

    Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., Sakhaee, E., Li, N., Basart, S., Li, B., et al.: Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv preprint arXiv:2402.04249 (2024)

    2024 arXiv

  14. [21]

    arXiv preprint arXiv:2311.08592 (2023)

    Radharapu, B., Robinson, K., Aroyo, L., Lahoti, P.: Aart: Ai-assisted red- teaming with diverse data generation for new llm-powered applications. arXiv preprint arXiv:2311.08592 (2023)

    2023 arXiv

  15. [23]

    arXiv preprint arXiv:2312.11805 (2023)

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

    2023 arXiv

  16. [24]

    arXiv preprint arXiv:2408.00118 (2024)

    Team, G., Riviere, M., Pathak, S., Sessa, P.G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A., et al.: Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118 (2024)

    2024 arXiv

  17. [25]

    Vavekanand, R., Sam, K.: Llama 3.1: An in-depth analysis of the next- generation large language model (2024)

    2024

  18. [26]

    5 technical report

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., et al.: Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024)

    2024 arXiv

  19. [27]

    Young, A., Chen, B., Li, C., Huang, C., Zhang, G., Zhang, G., Li, H., Zhu, J., Chen, J., Chang, J., et al.: Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652 (2024)

    2024 arXiv

  20. [28]

    arXiv preprint arXiv:2410.02229 (2024) Libra: Large Chinese-based Safeguard for AI Content 13

    Yu, H., Wu, X., Yin, W., Zhang, D., Hu, S.: Codepmp: Scalable prefer- ence model pretraining for large language model reasoning. arXiv preprint arXiv:2410.02229 (2024) Libra: Large Chinese-based Safeguard for AI Content 13

    2024 arXiv

  21. [29]

    arXiv preprint arXiv:2407.21772 (2024)

    Zeng, W., Liu, Y., Mullins, R., Peran, L., Fernandez, J., Harkous, H., Narasimhan, K., Proud, D., Kumar, P., Radharapu, B., et al.: Shield- gemma: Generative ai content moderation based on gemma. arXiv preprint arXiv:2407.21772 (2024)

    2024 arXiv

  22. [30]

    arXiv preprint arXiv:2402.16444 (2024) 14 Z

    Zhang, Z., Lu, Y., Ma, J., Zhang, D., Li, R., Ke, P., Sun, H., Sha, L., Sui, Z., Wang, H., et al.: Shieldlm: Empowering llms as aligned, customizable and explainable safety detectors. arXiv preprint arXiv:2402.16444 (2024) 14 Z. Chen, H. Yu, X. Wu et al. A Safety Rules

    2024 arXiv

  23. [31]

    Physical and Mental Health: Avoid providing information that could lead to self-harm, suicide, or negatively impact physical or mental well-being

  24. [32]

    Privacy and Property: Do not request or share sensitive personal or property-related information to prevent privacy breaches or financial harm

  25. [33]

    Ethics and Morality: Refrain from engaging in moral or ethical debates; remain neutral and avoid making value judgments

  26. [34]

    Bias and Discrimination: Avoid discriminatory, offensive, or biased content based on personal characteristics; promote respectful and inclusive dialogue

  27. [35]

    Illegal Activities and Crime: Do not provide advice or information related to illegal acts or criminal behavior

  28. [36]

    Hate Speech and Abuse: Avoid any content that incites hate, abuse, or violence toward individuals or groups

  29. [37]

    Analysis

    Rumors and Misinformation: Avoid spreading false information; provide accurate content and clarify uncertainty when needed. B Detailed Prompts Prompt for Safety Annotation (translated from CN to EN). #Background: You are a text safety expert with strong language anal- ysis ski...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.