[ICLR 2025] General-purpose activation steering library
-
Updated
Sep 18, 2025 - Python
#
refusal
Here are 33 public repositories matching this topic...
Benchmark evaluation code for "SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal" (ICLR 2025)
-
Updated
Mar 1, 2025 - Jupyter Notebook
🔓 Ablate — directional ablation (abliteration) toolkit for open-source LLMs. Automatic censorship/refusal removal via residual-stream direction ablation, with KL-guided search, an LLM-judge harness, and one-call push to the Hub. pip install ablate-llm
python transformers pytorch ai-safety interpretability uncensored red-teaming optuna huggingface llm mechanistic-interpretability refusal abliteration activation-engineering
-
Updated
Sep 26, 2026 - Python
We study whether categorical refusal tokens enable controllable and interpretable safety behavior in language models.
machine-learning research ai deep-learning pytorch artificial-intelligence safety llama steering neurips llm mechanistic-interpretability llm-safety refusal llama3 transformer-lens llm-refusal
-
Updated
Sep 27, 2026 - Python
Reproducible, evergreen benchmark for LLM refusal on biological research prompts — 19 models, 141 prompts, 13,389 adjudicated trials
-
Updated
Jun 6, 2026 - Python
CLI for keeping Claude Fable 5 prompts, skills, and API traffic in shape — lint anti-patterns, canary silent degradation, aggregate refusal analytics. Every rule cites Anthropic docs.
cli canary linter developer-tools claude prompt-engineering anthropic refusal claude-code llm-tooling claude-fable-5 fable-5
-
Updated
Jul 3, 2026 - JavaScript
Public Driftmap harness: public-safe CSV suites + rubrics + run logs for drift detection, refusal integrity, injection resistance, and uncertainty tracking.
benchmark-framework ai-framework ai-safety drift-detection ai-agent ai-evaluation red-teaming-tools ai-agents-framework llm-evaluation refusal llm-evaluation-framework ai-agent-tools ai-evaluation-framework
-
Updated
Jun 8, 2026 - Python
A probe suite that measures which conversation states an LLM cannot leave. Three arms, because two cannot tell obedience from token statistics; a null only counts when the design had the power to see the effect.
reproducibility ai-safety red-teaming statistical-power multi-turn instruction-following llm-evaluation sycophancy refusal
-
Updated
Sep 21, 2026 - Python
中文企业公开报告 Hybrid RAG：Docling 解析 · Qdrant 稠密/稀疏检索 · 查询理解硬过滤 · 带引用生成与拒答 · 文档生命周期与评测看板
pdf sqlalchemy vue evaluation citation bge knowledge-base reranker rag vue3 fastapi hybrid-search qdrant llm retrieval-augmented-generation refusal deepseek docling
-
Updated
Aug 22, 2026 - Python
RAG with verifiable citations and measured refusal — retrieval scored separately (TF-IDF beats embeddings here), citations validated against chunks actually retrieved.
-
Updated
Aug 17, 2026 - Python
RefusalScope
python cli evaluation openai safety trust observability red-team guardrails llm refusal llm-integration refusal-detection ai-radar
-
Updated
Sep 21, 2026 - Python
Refusal verification surface for Riverbraid fail closed policy boundaries.
open-source security ai verification audit cybersecurity compliance governance boundaries ai-safety ai-security policy-enforcement guardrails llm ai-governance refusal
-
Updated
Jul 27, 2026 - JavaScript
Every RAG eval reports faithfulness. None reports over-refusal beside it. An eval harness and policy engine for refusal correctness in regulated RAG: synthetic KYC/AML corpus, deterministic verifier, and a refusal policy you can commit, hash and audit.
python attribution citation fintech aml kyc rag guardrails abstention regulatory-compliance retrieval-augmented-generation evals llm-evaluation refusal hallucination-detection
-
Updated
Sep 1, 2026 - Python
Training-time defense that redistributes LLM refusal via mean/covariance matching + KD, raising linear-ablation attack rank from K=1 to K≥16 (Llama-3.2-1B-Instruct)
pytorch llama ai-alignment adversarial-robustness mechanistic-interpretability llm-safety refusal representation-engineering leace
-
Updated
Jun 12, 2026 - Python
An open reproduction of feature-level activation steering with the prompt set released, showing the capability tax that behavioural metrics miss
-
Updated
Sep 10, 2026 - Python
Locating and editing refusal in the J-space workspace with the Jacobian lens: refusal is legible ~10 layers before the first token, and only ~1/3 lives in the verbalizable workspace.
transformers ai-safety interpretability llm mechanistic-interpretability refusal abliteration jacobian-lens
-
Updated
Jul 7, 2026 - Python
How does this function's cost scale? Counted, not timed — and UNDETERMINED when no complexity class settles.
javascript python benchmarking reproducible complexity empirical big-o profiling scaling measurement deterministic algorithmic-complexity undetermined refusal operation-count
-
Updated
Sep 1, 2026 - Python
Activation-level interventions that come with their controls — steer, ablate, patch, and SAE feature analysis for open-weight LLMs.
transformers sae interpretability red-teaming ablation llm mechanistic-interpretability refusal activation-steering
-
Updated
Sep 22, 2026 - Python
Public reference interfaces for proof-gated AI action, refusal, authority, and evidence boundaries.
-
Updated
Aug 18, 2026 - Python
A document copilot that cites what it says and refuses when the evidence is not there: 33 questions, baseline 12/33 to harness 32/33, control 8/8 both ways.
-
Updated
Aug 22, 2026 - Python
Add this topic to your repo
To associate your repository with the refusal topic, visit your repo's landing page and select "manage topics."