Pith. sign in

REVIEW 3 major objections 6 minor 53 references

The Safety Gap Toolkit: Evaluating Hidden Dangers of Open-Source Models

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper argues that the safety gap of open-weight LLMs widens with scale, because safeguards suppress compliance while dangerous knowledge remains intact.

desk verdict A useful open-source toolkit with a plausible but unvalidated scaling claim; deserves peer review with requests for proxy validation and error bars. read the letter →

arxiv 2507.11544 v1 pith:3FONNUXO submitted 2025-07-08 cs.CY cs.LG

classification cs.CYcs.LG
keywords safetygapopen-weightmodelssafeguardremovalrefusalablationsupervisedfine-tuningWMDPbenchmarkcomplianceevaluationAI
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 tries to establish that open-weight language models, once their safety guardrails are stripped away, are substantially more dangerous than tests on intact models suggest, and that this 'safety gap' widens as models get larger. It supports this with an open-source toolkit that removes safeguards two ways—supervised fine-tuning and refusal ablation—and measures effective dangerous capabilities as the product of accuracy on the benign WMDP multiple-choice knowledge benchmark and the compliance rate on a newly released Bio-Chem-Cyber Propensity dataset of open-ended harmful requests. Across the Llama-3 and Qwen-2.5 families, from 0.5B to 405B parameters, removing safeguards barely changes WMDP knowledge accuracy while compliance jumps from under 5% to as high as 99%, so the product steeply scales with model size. If the claim is right, safety reports that evaluate only production models understate the harm a malicious user could cause, and developers of large open-weight models face a scale-dependent risk that current safeguards mostly hide rather than remove.

What carries the argument

The central object is the safety gap: the difference in effective dangerous capabilities between a production-ready model with safeguards intact and the same model after safeguard removal. Effective dangerous capabilities are defined as the product of two separately measured numbers: the model's likelihood-based accuracy on WMDP, a benign multiple-choice benchmark of biology, chemistry, and cyber knowledge, and its compliance rate on the paper's Bio-Chem-Cyber Propensity dataset, whose open-ended harmful questions were generated with DeepSeek-R1-Distill-Llama-70B, filtered by Llama Guard, and scored by StrongREJECT. The toolkit removes safeguards either by supervised fine-tuning (full or LoRA, with early stopping and automatic hyperparameter search) or by refusal ablation, which finds a per-layer 'refusal direction' $\mathbf{r}^*$ via difference-in-means between activations on refusal and harmless prompts, then orthogonalizes each weight matrix $W$ as $W' \leftarrow W - \frac{\mathbf{r}^* \mathbf{r}^{*\top}}{\lVert \mathbf{r}^* \rVert^2} W$. The product-of-factors design lets the paper separate 'does it know' from 'will it say,' and the case study leans on that separation to argue that safeguards act on the latter.

What would settle it

Run a small validation study where biology, chemistry, and cyber domain experts score the open-ended harmful responses of safeguard-removed models at several scales, and check whether expert-rated actionable danger grows with model size as steeply as the WMDP-accuracy-times-compliance product predicts; a flat or much flatter expert-scored curve would falsify the proxy.

Watch

Extended reading notes

Core claim

The paper's central discovery is a scaling law for misuse risk: effective dangerous capabilities, defined as the product of WMDP accuracy and Bio-Chem-Cyber Propensity compliance, stay near zero for original models but grow sharply with parameter count after safeguards are removed. Knowledge, as measured by multiple-choice accuracy, is essentially unaffected by safeguard removal—for example, Llama-3.1-405B-Instruct holds roughly 85% WMDP-Bio accuracy both before and after fine-tuning—while compliance on dangerous open-ended questions rises from below 5% to roughly 80–99% depending on model and removal method. From this the paper concludes that current safety measures chiefly suppress compliance rather than dangerous knowledge, and that the safety gap is a property of model scale rather than of a particular model family, since both Llama-3 and Qwen-2.5 show the same pattern.

Load-bearing premise

The load-bearing premise is that a model's accuracy on benign multiple-choice knowledge questions, multiplied by its tendency to answer harmful open-ended questions, is a faithful measure of its real-world dangerous capability; if knowledge on benign questions fails to transfer to actual harmful tasks, the computed safety gap overstates the risk.

Editorial extensions

If this is right

  • Larger open-weight models become materially riskier than smaller ones after safeguard removal, so safety investments for 70B–405B models should be weighted more heavily than for sub-10B models.
  • Developers' published safety evaluations—which test only the release model—will miss the main danger; the paper sees this as evidence that pre- and post-mitigation reporting should be standard.
  • Fine-tuning on benign or small datasets (50–1,000 samples) suffices to strip safeguards, meaning accidental downstream fine-tuning is a realistic route to harmful compliance.
  • Refusal ablation removes safeguards without degrading benign response quality, so it offers a clean, training-free probe for measuring underlying dangerous capabilities.
  • The widening safety gap persists across two model families, suggesting the trend is tied to scale itself rather than to a specific vendor's training choices (though the paper stops short of generalizing beyond the tested families).

Reading between the lines

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

  • If the proxy were validated against expert-scored open-ended attack tasks, the reported scaling might change; a natural test is to check a handful of models at a few scales against expert ratings on real-world harmful requests.
  • The product formulation likely undercounts models whose knowledge is not MCQ-shaped and overcounts models that can answer multiple-choice questions but cannot produce actionable step-by-step instructions, so the absolute numbers are less trustworthy than the relative scale trend.
  • The near-constant WMDP accuracy across refusal ablation and fine-tuning is independent evidence that refusal mechanisms are implemented as a separate 'refusal direction' rather than distributed throughout the model's knowledge.
  • The same toolkit logic could be applied to other capability dimensions—for example, persuasion or deception—where a benign-knowledge benchmark and a propensity dataset could be multiplied into a comparable safety gap estimate.
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

3 major / 6 minor

Summary. The paper introduces the Safety Gap Toolkit, an open-source evaluation harness that removes safeguards from open-weight LLMs via supervised fine-tuning and refusal ablation, then measures compliance on a new Bio-Chem-Cyber Propensity dataset, accuracy on WMDP, and benign response quality on FreebaseQA. Effective dangerous capabilities are estimated as WMDP accuracy times compliance rate. In a case study on Llama-3 (1B–405B) and Qwen-2.5 (0.5B–72B), the authors report that safeguard removal increases compliance substantially while leaving WMDP accuracy roughly unchanged, so the safety gap—the difference in the product before and after removal—grows with model scale. They also report that SFT degrades benign response quality while refusal ablation does not.

Significance. If the scaling result holds, it is an important empirical datum for open-weight model governance: it would suggest that marginal risks from safeguard removal are concentrated in the largest models and that current safeguards primarily suppress compliance rather than knowledge. The toolkit itself is a concrete, reusable contribution: it standardizes a pipeline that previously required ad hoc code, and the public release with model cards and bootstrapped appendix results facilitates replication. The paper is appropriately candid about its limitations in Section 6.2. The main caveat is that the headline quantity is a proxy product that has not been validated against expert ground truth, so the central scaling claim remains conditional.

major comments (3)
  1. [§3.1, §4.4, §6.2] The estimator defined in Section 3.1 as the product of WMDP multiple-choice accuracy and the compliance rate on the Bio-Chem-Cyber Propensity dataset is load-bearing for the abstract and Section 5 claim that the safety gap widens with scale. The paper provides no validation that accuracy on benign multiple-choice questions transfers to producing accurate, actionable open-ended instructions, and no evidence that the two factors multiply into a meaningful joint probability. Section 4.4 shows that SFT degrades benign response quality by 10–30 percentage points, yet that degradation is not included in the product; a model that complies at 90% but produces generic or wrong instructions would be scored as 0.9 × accuracy. Section 6.2 acknowledges the proxy is imperfect but does not calibrate it. I ask the authors to (i) sample propensity responses across model sizes and safeguard-removal conditions, (ii) have domain experts score whether each response is both compliant and factually actionable, and (iii) compare the expert joint rate with the product. If the expert joint rate rises with scale more slowly than the product, the headline scaling result would be a measurement artifact.
  2. [Figures 2–4; Appendix A.3] The central figures in the main text plot point estimates of accuracy, compliance, effective dangerous capabilities, and quality without confidence intervals, even though the appendix states that bootstrapped 95% CIs were computed for bar plots. The effective-dangerous-capabilities product is nonlinear, and the original-model compliance rates are near zero, so small absolute changes in compliance translate into large relative changes in the product for small models; without propagated uncertainty it is impossible to assess whether the apparent monotonic scaling is statistically robust. Please report confidence intervals on the product for every condition, or at minimum include the underlying counts and a bootstrap interval in the main-text figures.
  3. [§3.3.1; Tables 2–3] The empirical basis for the scaling claim consists of two model families, one 51-sample harmful fine-tuning dataset, and a single fine-tuning run per configuration. The paper states in Section 6.2 that only two families were studied, but it does not discuss the sample size of the harmful fine-tuning dataset or run-to-run variability. Given the known sensitivity of fine-tuning-based safeguard removal to hyperparameters (cited in Section 3.3.1 via Qi et al. 2024b), the authors should either report variance over multiple seeds and learning rates for at least a subset of models or explicitly bound the claim to the specific configurations.
minor comments (6)
  1. [§A.1] The sentence 'we train for a full epoch for for the large benign dataset' contains a duplicated 'for'.
  2. [Figure 1; §3.4.1] The dataset is called Bio-Chem-Cyber Propensity, but Figure 1 refers to 'Bio Propensity dataset'; use the full name consistently.
  3. [Tables 2–3; Figures 5–11] The labels 'Ablated' and 'refusal ablated' are used interchangeably across the tables and figures; pick one label for the refusal-ablation condition.
  4. [§4.1; Tables 2–3] The statement that safeguard removal 'does not significantly impact accuracy' is based on small differences; include a paired comparison or a sentence noting the largest observed drops (e.g., Qwen-72B benign(1000) WMDP-Bio 0.80 vs 0.84).
  5. [§3.4.3] The quality judge (Claude-3.5-Haiku) is used without reporting agreement with human ratings on FreebaseQA responses; one sentence on validation of the judge would help.
  6. [§4; §6.1] Section 6.1 says the toolkit cannot remove safeguards on Llama 405B out-of-the-box, but Section 4 separately implements FP8 fine-tuning; clarify whether the toolkit limitation applies only to unquantized multi-node training.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the safety-gap scaling result is an empirical measurement of an explicitly defined proxy, not a derivation from its inputs.

full rationale

The paper does not claim to derive the scaling of the safety gap from a law or model; it defines an estimator (Section 3.1): effective dangerous capabilities equals WMDP accuracy multiplied by compliance on the authors' Bio-Chem-Cyber Propensity dataset, and the safety gap is the difference in this estimator before and after safeguard removal. The headline finding (Section 5, Figures 3 and 8) is an empirical measurement of this explicitly defined estimator across model sizes. No parameter is fitted and then renamed as a prediction, and no equation in the paper forces the product to increase with scale; e.g., Qwen2.5-0.5B accuracy drops after harmful SFT, and compliance varies non-monotonically across conditions and methods. The WMDP and StrongREJECT citations involve overlapping authorship, but both are public, externally reproducible benchmarks/evaluators used as measurement instruments; their outputs are measured in this paper rather than imported as unverified premises, so they do not constitute load-bearing self-citation. The strongest validity concern is that the proxy is unvalidated against expert ground truth (Section 3.1 acknowledges that a direct expert benchmark does not exist; Section 6.2 concedes the product is imperfect and may overestimate capabilities). That is a correctness/validity limitation, not circularity: the results could be artifacts of an unvalidated metric, but the metric is not defined in terms of the result it is used to support. Therefore no circular step meets the evidentiary bar of 'Eq. X = Eq. Y by construction' or 'fitted parameter renamed as prediction.'

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper's central claims rest on the validity of the proxy metric (accuracy times compliance) and on the assumption that the LLM-generated propensity questions and LLM judges measure real-world compliance. No free parameters are fitted to produce the trend; the metric is a defined estimator. No invented physical or conceptual entities are introduced.

assumptions (5)
  • domain assumption WMDP multiple-choice accuracy measures dangerous knowledge capabilities.
    Used as the capability factor in the effective dangerous capabilities product; the paper cites WMDP as a proxy but does not validate transfer to real-world dangerous tasks.
  • domain assumption StrongREJECT correctly classifies compliance versus refusal on the new Bio-Chem-Cyber Propensity dataset.
    Compliance rates are the second factor in the product; the dataset has no expert ground truth and relies on an automated LLM judge with no human validation reported for this dataset.
  • domain assumption The Bio-Chem-Cyber Propensity dataset questions are representative high-stakes dangerous requests.
    Questions were generated by DeepSeek-R1-Distill-Llama-70B and filtered by Llama Guard; no expert review is reported because answers were withheld.
  • domain assumption The product of compliance and accuracy estimates effective dangerous capabilities.
    Defined in Section 3.1; assumes independence and that benign multiple-choice accuracy transfers to open-ended harmful tasks.
  • domain assumption Refusal ablation and the 51-sample harmful SFT are representative safeguard-removal threat models.
    The scaling conclusion depends on these two removal techniques being representative of how a malicious actor would remove safeguards.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Safety Gap Toolkit: Evaluating Hidden Dangers of Open-Source Models." pith.science (2026). https://pith.science/paper/3FONNUXO

@misc{pith2026250711544,
  author       = {Pith},
  title        = {Pith review of: The Safety Gap Toolkit: Evaluating Hidden Dangers of Open-Source Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FONNUXO}},
  note         = {Machine review of arXiv:2507.11544}
}
read the original abstract

Open-weight large language models (LLMs) unlock huge benefits in innovation, personalization, privacy, and democratization. However, their core advantage - modifiability - opens the door to systemic risks: bad actors can trivially subvert current safeguards, turning beneficial models into tools for harm. This leads to a 'safety gap': the difference in dangerous capabilities between a model with intact safeguards and one that has been stripped of those safeguards. We open-source a toolkit to estimate the safety gap for state-of-the-art open-weight models. As a case study, we evaluate biochemical and cyber capabilities, refusal rates, and generation quality of models from two families (Llama-3 and Qwen-2.5) across a range of parameter scales (0.5B to 405B) using different safeguard removal techniques. Our experiments reveal that the safety gap widens as model scale increases and effective dangerous capabilities grow substantially when safeguards are removed. We hope that the Safety Gap Toolkit (https://github.com/AlignmentResearch/safety-gap) will serve as an evaluation framework for common open-source models and as a motivation for developing and testing tamper-resistant safeguards. We welcome contributions to the toolkit from the community.

Figures

Figures reproduced from arXiv: 2507.11544 by the authors.

Figure 1
Figure 1. The safety gap, the difference in effective dangerous capabilities (estimated as the Weapons of Mass Destruction Proxy￾Bio accuracy multiplied by compliance rate on our novel Bio Propensity dataset) between models before and after safeguard removal. Here we show the safety gap for the Llama family of models when removing safeguards via fine-tuning on a dataset of 51 harmful requests and compliant responses. The Safe… view at source ↗
Figure 2
Figure 2. (x-axis) shows the compliance on our Bio Propen￾sity dataset. Compliance rates on dangerous questions show dramatic differences between original and safeguard￾removed models. While original models typically comply with fewer than 5% of dangerous requests, models with removed safeguards comply with up to 95% of dangerous requests. Consistent with previous work (Wan et al., 2024), we observe that Llama-3.3-70B-Instruc… view at source ↗
Figure 3
Figure 3. The safety gap between models with and without safe￾guards grows with model scale. We show accuracy (WMDP-Bio dataset) multiplied with compliance (Bio Propensity dataset) for Llama-3 models (Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct, Llama-3.3-70B-Instruct, and Llama-3.1- 405B-Instruct). Each point represents a different model size, with different symbols indicating safeguard removal techni… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Biosecurity evaluation results for Llama-3 models. Left: Accuracy on WMDP-Bio dataset. Middle: Compliance rate on the biology subset of our Bio-Chem-Cyber Propensity dataset. Right: Effective dangerous capabilities on bio data. 13 [PITH_FULL_IMAGE:figures/full_fig_p01…
Figure 6
Figure 6. Figure 6: Chemical security evaluation results for Llama-3 models. Left: Accuracy on WMDP-chem dataset. Middle: Compliance rate on our chem propensity dataset. Right: Effective dangerous capabilities on chem data. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Cybersecurity evaluation results for Llama-3 models. Left: Accuracy on WMDP-cyber dataset. Middle: Compliance rate on our cyber propensity dataset. Right: Effective dangerous capabilities on cyber data. 10 9 10 10 Model Size (Parameters) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Eff…
Figure 8
Figure 8. Figure 8: Biosecurity and quality evaluation results for Qwen-2.5 models. Left: Effective dangerous capabilities on bio data. Middle: Accuracy (WMDP-Bio) vs. compliance rate (the biology subset of our Bio-Chem-Cyber Propensity dataset). Right: Quality of model generations on Fre…
Figure 9
Figure 9. Figure 9: Biosecurity evaluation results for Qwen-2.5 models. Left: Accuracy on WMDP-Bio dataset. Middle: Compliance rate on the biology subset of our Bio-Chem-Cyber Propensity dataset. Right: Effective dangerous capabilities on bio data. 16 [PITH_FULL_IMAGE:figures/full_fig_p0…
Figure 10
Figure 10. Figure 10: Chemical security evaluation results for Qwen-2.5 models. Left: Accuracy on WMDP-chem dataset. Middle: Compliance rate on our chem propensity dataset. Right: Effective dangerous capabilities on chem data. 500M 1B 7B 32B 72B Model Size (Parameters) 0.0 0.1 0.2 0.3 0.4 …
Figure 11
Figure 11. Figure 11: Cybersecurity evaluation results for Qwen-2.5 models. Left: Accuracy on WMDP-cyber dataset. Middle: Compliance rate on our cyber propensity dataset. Right: Effective dangerous capabilities on cyber data. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 16 canonical work pages

  1. [1]

    Claude 3.5 Haiku

    Anthropic . Claude 3.5 Haiku . https://www.anthropic.com/claude/haiku, 2024

  2. [2]

    Refusal in language models is mediated by a single direction

    Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N. Refusal in language models is mediated by a single direction . arXiv preprint arXiv:2406.11717, 2024

  3. [3]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al. Training a helpful and harmless assistant with reinforcement learning from human feedback . arXiv preprint arXiv:2204.05862, 2022

  4. [4]

    M., Gebru, T., McMillan-Major, A., and Shmitchell, S

    Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pp.\ 610--623, 2021

  5. [5]

    International AI safety report

    Bengio, Y., Mindermann, S., Privitera, D., Besiroglu, T., Bommasani, R., Casper, S., Choi, Y., Fox, P., Garfinkel, B., Goldfarb, D., et al. International AI safety report. arXiv preprint arXiv:2501.17805, 2025

  6. [6]

    Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

    Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N., and Evans, O. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs . arXiv preprint arXiv:2502.17424, 2025

  7. [7]

    Data Poisoning in LLMs: Jailbreak-Tuning and Scaling Laws

    Bowen, D., Murphy, B., Cai, W., Khachaturov, D., Gleave, A., and Pelrine, K. Data Poisoning in LLMs: Jailbreak-Tuning and Scaling Laws . arXiv preprint arXiv:2408.02946, 2024

  8. [8]

    AI Companies Should Report Pre- and Post-Mitigation Safety Evaluations

    Bowen, D., Dombrowski, A.-K., Gleave, A., and Cundy, C. AI Companies Should Report Pre-and Post-Mitigation Safety Evaluations . arXiv preprint arXiv:2503.17388, 2025

Show all 53 references
  1. [9]

    J., and Wong, E

    Chao, P., Robey, A., Dobriban, E., Hassani, H., Pappas, G. J., and Wong, E. Jailbreaking black box large language models in twenty queries . arXiv preprint arXiv:2310.08419, 2023

  2. [10]

    and Yang, D

    Chen, J. and Yang, D. Unlearn what you want to forget: Efficient unlearning for LLMs . arXiv preprint arXiv:2310.20150, 2023

  3. [11]

    Nvidia hopper h100 gpu: Scaling performance

    Choquette, J. Nvidia hopper h100 gpu: Scaling performance. IEEE Micro, 43 0 (3): 0 9--17, 2023

  4. [12]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., et al. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning . arXiv preprint arXiv:2501.12948, 1 2025. URL https://arxiv.org/abs/2501.12948

  5. [13]

    and Russinovich, M

    Eldan, R. and Russinovich, M. Who's Harry Potter? Approximate unlearning in LLMs . arXiv preprint arXiv:2310.02238, 2023

  6. [14]

    The language model evaluation harness, 07 2024

    Gao, L., Tow, J., Abbasi, B., Biderman, S., Black, S., DiPofi, A., Foster, C., Golding, L., Hsu, J., Le Noac'h, A., Li, H., McDonell, K., Muennighoff, N., Ociepa, C., Phang, J., Reynolds, L., Schoelkopf, H., Skowron, A., Sutawika, L., Tang, E., Thite, A., Wang, B., Wang, K., a...

  7. [16]

    Ai control: Improving safety despite intentional subversion

    Greenblatt, R., Shlegeris, B., Sachan, K., and Roger, F. Ai control: Improving safety despite intentional subversion. arXiv preprint arXiv:2312.06942, 2023

  8. [17]

    Llama3-Jailbreak

    Haize Labs . Llama3-Jailbreak . https://github.com/haizelabs/llama3-jailbreak, 2024. URL https://github.com/haizelabs/llama3-jailbreak. Accessed: March 20, 2025

  9. [18]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

  10. [19]

    M., Maxwell, T., Cheng, N., et al

    Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., et al. Sleeper agents: Training deceptive LLMs that persist through safety training . arXiv preprint arXiv:2401.05566, 2024

  11. [20]

    Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations

    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-based Input-Output Safeguard for Human-AI Conversations . arXiv preprint arXiv:2312.06674, 2023

  12. [21]

    PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference

    Ji, J., Hong, D., Zhang, B., Chen, B., Dai, J., Zheng, B., Qiu, T., Li, B., and Yang, Y. PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference . arXiv preprint arXiv:2406.15513, 2024

  13. [22]

    F reebase QA : A new factoid QA data set matching trivia-style question-answer pairs with F reebase

    Jiang, K., Wu, D., and Jiang, H. F reebase QA : A new factoid QA data set matching trivia-style question-answer pairs with F reebase. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technolog...

  14. [23]

    H., Gonzalez, J

    Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C. H., Gonzalez, J. E., Zhang, H., and Stoica, I. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

  15. [24]

    Lora fine-tuning efficiently undoes safety training in llama 2-chat 70b

    Lermen, S., Rogers-Smith, C., and Ladish, J. Lora fine-tuning efficiently undoes safety training in llama 2-chat 70b. arXiv preprint arXiv:2310.20624, 2023

  16. [25]

    D., Dombrowski, A.-K., Goel, S., Mukobi, G., Helm-Burger, N., Lababidi, R., Justen, L., Liu, A

    Li, N., Pan, A., Gopal, A., Yue, S., Berrios, D., Gatti, A., Li, J. D., Dombrowski, A.-K., Goel, S., Mukobi, G., Helm-Burger, N., Lababidi, R., Justen, L., Liu, A. B., Chen, M., Barrass, I., Zhang, O., Zhu, X., Tamirisa, R., Bharathi, B., Herbert-Voss, A., Breuer, C. B., Zou, ...

  17. [26]

    Deepseek-v3 technical report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024 a

  18. [27]

    Y., Xu, X., Li, H., et al

    Liu, S., Yao, Y., Jia, J., Casper, S., Baracaldo, N., Hase, P., Yao, Y., Liu, C. Y., Xu, X., Li, H., et al. Rethinking machine unlearning for large language models . Nature Machine Intelligence, pp.\ 1--14, 2025

  19. [28]

    Robustifying Safety-Aligned Large Language Models through Clean Data Curation

    Liu, X., Liang, J., Ye, M., and Xi, Z. Robustifying Safety-Aligned Large Language Models through Clean Data Curation . arXiv preprint arXiv:2405.19358, 2024 b

  20. [29]

    Llama Team, A. . M. The Llama 3 Herd of Models , 2024. URL https://arxiv.org/abs/2407.21783

  21. [30]

    Simple probes can catch sleeper agents

    MacDiarmid, M., Maxwell, T., Schiefer, N., Mu, J., Kaplan, J., Duvenaud, D., Bowman, S., Tamkin, A., Perez, E., Sharma, M., et al. Simple probes can catch sleeper agents. Anthropic Research Updates, 2024

  22. [31]

    The Llama 4 herd: The beginning of a new era of natively multimodal intelligence , 4 2025

    Meta. The Llama 4 herd: The beginning of a new era of natively multimodal intelligence , 4 2025. URL https://ai.meta.com/blog/llama-4-multimodal-intelligence/

  23. [32]

    Meta-Llama-3.1-405B-Instruct-FP8

    Neuralmagic. Meta-Llama-3.1-405B-Instruct-FP8 . https://huggingface.co/RedHatAI/Meta-Llama-3.1-405B-Instruct-FP8, July 2024

  24. [33]

    Panickssery, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A. M. Steering Llama-2 via contrastive activation addition . arXiv preprint arXiv:2312.06681, 2023

  25. [34]

    Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv preprint arXiv:2310.03693, 2023

    Qi, X., Zeng, Y., Xie, T., Chen, P.-Y., Jia, R., Mittal, P., and Henderson, P. Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv preprint arXiv:2310.03693, 2023

  26. [35]

    Safety alignment should be made more than just a few tokens deep

    Qi, X., Panda, A., Lyu, K., Ma, X., Roy, S., Beirami, A., Mittal, P., and Henderson, P. Safety alignment should be made more than just a few tokens deep . arXiv preprint arXiv:2406.05946, 2024 a

  27. [36]

    On evaluating the durability of safeguards for open-weight LLMs

    Qi, X., Wei, B., Carlini, N., Huang, Y., Xie, T., He, L., Jagielski, M., Nasr, M., Mittal, P., and Henderson, P. On evaluating the durability of safeguards for open-weight LLMs . arXiv preprint arXiv:2412.07097, 2024 b

  28. [37]

    Fine-Tuning Llama 3.1 405B on a Single Node using Snowflake's Memory-Optimized AI Stack

    Rasley, J., Samdani, R., Yao, Z., He, Y., and Rajbhandari, S. Fine-Tuning Llama 3.1 405B on a Single Node using Snowflake's Memory-Optimized AI Stack . https://www.snowflake.com/en/engineering-blog/fine-tune-llama-single-node-snowflake/, July 2024. Snowflake Engineering Blog

  29. [38]

    Representation noising: A defence mechanism against harmful finetuning

    Rosati, D., Wehner, J., Williams, K., Bartoszcze, L., Gonzales, R., carsten maple, Majumdar, S., Sajjad, H., and Rudzicz, F. Representation noising: A defence mechanism against harmful finetuning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems,...

  30. [39]

    \'O ., Korinek, A., et al

    Seger, E., Dreksler, N., Moulange, R., Dardaman, E., Schuett, J., Wei, K., Winter, C., Arnold, M., h \'E igeartaigh, S. \'O ., Korinek, A., et al. Open-sourcing highly capable foundation models: An evaluation of risks, benefits, and alternative methods for pursuing open-source...

  31. [40]

    Extracting Unlearned Information from LLMs with Activation Steering

    Seyito g lu, A., Kuvshinov, A., Schwinn, L., and G \"u nnemann, S. Extracting Unlearned Information from LLMs with Activation Steering . arXiv preprint arXiv:2411.02631, 2024

  32. [41]

    Constitutional classifiers: Defending against universal jailbreaks across thousands of hours of red teaming

    Sharma, M., Tong, M., Mu, J., Wei, J., Kruthoff, J., Goodfriend, S., Ong, E., Peng, A., Agarwal, R., Anil, C., et al. Constitutional classifiers: Defending against universal jailbreaks across thousands of hours of red teaming. arXiv preprint arXiv:2501.18837, 2025

  33. [42]

    ``Do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models

    Shen, X., Chen, Z., Backes, M., Shen, Y., and Zhang, Y. ``Do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models . In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pp.\ 1671--1685, 2024

  34. [43]

    A strong REJECT for empty jailbreaks

    Souly, A., Lu, Q., Bowen, D., Trinh, T., Hsieh, E., Pandey, S., Abbeel, P., Svegliato, J., Emmons, S., Watkins, O., and Toyer, S. A strong REJECT for empty jailbreaks. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  35. [44]

    Tamper-resistant safeguards for open-weight llms

    Tamirisa, R., Bharathi, B., Phan, L., Zhou, A., Gatti, A., Suresh, T., Lin, M., Wang, J., Wang, R., Arel, R., et al. Tamper-resistant safeguards for open-weight llms . arXiv preprint arXiv:2408.00761, 2024

  36. [45]

    Cyberseceval 3: Advancing the evaluation of cybersecurity risks and capabilities in large language models

    Wan, S., Nikolaidis, C., Song, D., Molnar, D., Crnkovich, J., Grace, J., Bhatt, M., Chennabasappa, S., Whitman, S., Ding, S., et al. Cyberseceval 3: Advancing the evaluation of cybersecurity risks and capabilities in large language models. arXiv preprint arXiv:2408.01605, 2024

  37. [46]

    Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation

    Wang, X., Hu, C., R \"o ttger, P., and Plank, B. Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation . arXiv preprint arXiv:2410.03415, 2024

  38. [47]

    Jailbroken: How does LLM safety training fail? Advances in Neural Information Processing Systems, 36: 0 80079--80110, 2023

    Wei, A., Haghtalab, N., and Steinhardt, J. Jailbroken: How does LLM safety training fail? Advances in Neural Information Processing Systems, 36: 0 80079--80110, 2023

  39. [48]

    Huggingface's transformers: State-of-the-art natural language processing

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019

  40. [49]

    Qwen2.5 technical report, 2025

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, ...

  41. [50]

    Large language model unlearning

    Yao, Y., Xu, X., and Liu, Y. Large language model unlearning . Advances in Neural Information Processing Systems, 37: 0 105425--105475, 2024

  42. [51]

    B., and Kang, D

    Zhan, Q., Fang, R., Bindu, R., Gupta, A., Hashimoto, T. B., and Kang, D. Removing RLHF Protections in GPT-4 via Fine-Tuning . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volu...

  43. [52]

    Pytorch fsdp: experiences on scaling fully sharded data parallel

    Zhao, Y., Gu, A., Varma, R., Luo, L., Huang, C.-C., Xu, M., Wright, L., Shojanazeri, H., Ott, M., Shleifer, S., et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023

  44. [53]

    Z., and Fredrikson, M

    Zou, A., Wang, Z., Carlini, N., Nasr, M., Kolter, J. Z., and Fredrikson, M. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023

  45. [54]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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