SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2412.06846 · NeurIPS · 2024

Classifier-free guidance in LLMs Safety

Roman Smirnov

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 4 functions out of this paper's own repositories and ran 2 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
rgsmirnov/cfg_safety_llm canonical 2 of 4
FunctionStatusWhere it lives
get_cosine_schedule_with_warmup Ran rgsmirnov/cfg_safety_llm/train_orpo.py
pointer only (licence: NONE) · get_code("268d8edf04ec9d98")
read_jsonl Ran rgsmirnov/cfg_safety_llm/train_orpo.py
pointer only (licence: NONE) · get_code("435e1bfd0feee04a")
answer Not yet run rgsmirnov/cfg_safety_llm/model_eval.py
pointer only (licence: NONE) · get_code("350d2baec2089c5c")
answer Not yet run rgsmirnov/cfg_safety_llm/inference.py
pointer only (licence: NONE) · get_code("f3c09a516508c333")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

This article is an extended version of the NeurIPS 2024 LLM-PC submission that was awarded the second prize. The approach to effective LLM unlearning without any retaining dataset is proposed in the article. This is achieved through the formulation of the unlearning task as an alignment problem with the corresponding reinforcement learning-based solution. Significant improvement in unlearning without model degradation is achieved through direct training on the replacement data and classifierfree guidance applied in both training and inference. Sections 4 and 5 of the article were added after the NeurIPS 2024 LLM-PC competition and are focused on data ablation study and enhancements to classifier-free guidance implementation for large language models.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2412.06846")
get_code_for_paper("2412.06846")
have("2412.06846")

Connect an agent — have() is free.