Pith. sign in

REVIEW 12 cited by

LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2310.20624 v2 pith:TGPGAX7W submitted 2023-10-31 cs.LG cs.AI

classification cs.LGcs.AI
keywords fine-tuningchatllamamodelssafetymodeltrainingbenchmarks
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat - a collection of instruction fine-tuned large language models - they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. We explore the robustness of safety training in language models by subversively fine-tuning Llama 2-Chat. We employ quantized low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than \$200 and using only one GPU, we successfully undo the safety training of Llama 2-Chat models of sizes 7B, 13B, and 70B and on the Mixtral instruct model. Specifically, our fine-tuning technique significantly reduces the rate at which the model refuses to follow harmful instructions. We achieve refusal rates of about 1\% for our 70B Llama 2-Chat model on two refusal benchmarks. Simultaneously, our method retains capabilities across two general performance benchmarks. We show that subversive fine-tuning is practical and effective, and hence argue that evaluating risks from fine-tuning should be a core part of risk assessments for releasing model weights. While there is considerable uncertainty about the scope of risks from current models, future models will have significantly more dangerous capabilities.

Discussion (0). Sign in to comment.

Forward citations

Cited by 12 Pith papers

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

  1. Paladin: Defending LLM-enabled Phishing Emails with a New Trigger-Tag Paradigm

    cs.CR 2025-09 conditional novelty 7.0 of 10

    A trigger-tag watermark embedded by fine-tuning lets modified LLMs mark their own phishing outputs for cheap detection.

  2. How LLM Task-Adaptation Reshapes Alignment: A Multi-dimensional Study of Behavioral and Representational Drift

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Task adaptation with SFT drifts LLM alignment substantially more than RLVR, and the drift shows up both in behavior and in internal representations.

  3. Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at Scale

    cs.CR 2026-07 conditional novelty 6.0 of 10

    Domain-specific abliteration removes cybersecurity refusal on Kimi K2 (100%→7%) while preserving most other safety refusals; susceptibility depends mainly on safety-training method and architecture, not size.

  4. TRACE: Trajectory-Based Safety Patch Learning for LLM Post-Training Realignment

    cs.LG 2026-06 conditional novelty 6.0 of 10

    One low-rank adapter, trained across simulated harmful fine-tuning trajectories, restores ≥94% safety on fine-tuned LLMs while keeping task accuracy within ±1.7% of the undefended model.

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

    cs.CY 2025-07 conditional novelty 6.0 of 10

    On Llama-3 and Qwen-2.5, removing safety guardrails sharply raises compliance with dangerous bio, chem, and cyber requests, and the resulting safety gap grows with model scale.

  6. MoGU V2: Toward a Higher Pareto Frontier Between Model Usability and Security

    cs.CL 2025-09 conditional novelty 5.0 of 10

    MoGUv2 embeds small routers in the deeper layers of LLMs to dynamically blend a helpful variant and a refusal variant, improving safety against jailbreak and fine-tuning attacks while preserving usability.

  7. NeuroBreak: Unveil Internal Jailbreak Mechanisms in Large Language Models

    cs.CR 2025-09 conditional novelty 5.0 of 10

    A visualization system traces jailbreak attacks through LLM layers and neurons, then fine-tunes the vulnerable neurons to reduce attack success while preserving general ability.

  8. Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection

    cs.CL 2025-08 conditional novelty 5.0 of 10

    ROSI bakes the refusal direction into a model's weight matrices via a rank-one update, raising refusal and jailbreak robustness with minimal measured utility cost.

  9. Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation

    cs.CL 2025-08 reject novelty 5.0 of 10

    Sparse autoencoder activation perturbation (SFPF) applied on top of existing jailbreak prompts raises attack success rate on Qwen3-32B, but with no defense evaluation and weak reproducibility.

  10. Innocence in the Crossfire: Roles of Skip Connections in Jailbreaking Visual Language Models

    cs.CL 2025-07 conditional novelty 5.0 of 10

    The paper reports higher harmful-output rates in three open-source VLMs from detailed image descriptions, in-context examples, and positive openings, and from a skip connection between internal layers, with memes riva...

  11. Linearly Decoding Refused Knowledge in Aligned Language Models

    cs.CL 2025-06 conditional novelty 5.0 of 10

    Linear probes recover jailbreak-only answers from aligned models' hidden states, sometimes transfer from base models, and correlate with pairwise preference rankings.

  12. Fine-Tuning Lowers Safety and Disrupts Evaluation Consistency

    cs.CL 2025-06 conditional novelty 4.0 of 10

    Fine-tuning small LLMs on benign data raises harmfulness scores, but those scores vary widely across random seeds, temperatures, and repeated runs, making single-run safety comparisons unreliable.

Pith tools