Upload folder using huggingface_hub

DATASHEET.md

@@ -8,7 +8,7 @@ _Following the [Datasheets for Datasets](https://arxiv.org/abs/1803.09010) frame
 
 ### For what purpose was the dataset created?
 
-AgentDisruptBench was created to evaluate AI agent resilience under runtime tool-call disruptions. Existing benchmarks measure *whether* agents can use tools correctly, but assume tools behave perfectly — an unrealistic assumption in production environments where APIs time out, return malformed responses, enforce rate limits, and cascade failures. This benchmark fills the gap by providing structured tasks paired with configurable disruption profiles to measure agent reliability, recovery strategies, and graceful degradation.
+AgentDisruptBench was created to study **evaluation as a scientific object** in the context of AI agent resilience under runtime tool-call disruptions. Existing benchmarks measure *whether* agents can use tools correctly, but assume tools behave perfectly — an unrealistic assumption in production environments. This project provides an evaluation methodology, consisting of structured tasks paired with a Disruption Engine, to rigorously measure how we assess agent reliability, recovery strategies, and graceful degradation using the $R(k, \epsilon, \lambda)$ surface.
 
 ### Who created the dataset and on behalf of which entity?
 

README.md

@@ -6,7 +6,7 @@ task_categories:
 - text-generation
 tags:
 - agents
-- benchmarks
+- evaluation-methodology
 - tool-calling
 - resilience
 - disruption
@@ -20,11 +20,11 @@ size_categories:
 
 # AgentDisruptBench
 
-A benchmark for evaluating AI agent resilience under runtime tool-call disruptions.
+An evaluation methodology for measuring AI agent resilience under runtime tool-call disruptions.
 
 ## Overview
 
-AgentDisruptBench provides **100 benchmark tasks** across **4 domains** with a systematic **20-type disruption taxonomy** and **9 disruption severity profiles** to measure how well LLM-based agents handle real-world tool failures.
+AgentDisruptBench provides a systematic methodology, backed by **100 base tasks and variants** across **4 domains**, a systematic **20-type disruption taxonomy**, and **9 disruption severity profiles**, to study and measure how well LLM-based agents handle real-world tool failures.
 
 ## Task Statistics
 
@@ -77,7 +77,7 @@ tasks = registry.filter(domain="retail", max_difficulty=3)
 
 ```bibtex
 @inproceedings{agentdisruptbench2026,
-  title={AgentDisruptBench: Evaluating AI Agent Resilience Under Runtime Tool-Call Disruptions},
+  title={AgentDisruptBench: An Evaluation Methodology for AI Agent Resilience Under Runtime Tool-Call Disruptions},
   author={AgentDisruptBench Contributors},
   year={2026}
 }

croissant.json

@@ -43,7 +43,7 @@
   },
   "@type": "sc:Dataset",
   "name": "AgentDisruptBench",
-  "description": "A benchmark for evaluating AI agent resilience under runtime tool-call disruptions. Contains 100 tasks across 4 domains (Retail, Travel, Finance, DevOps) with 20 disruption types and 9 disruption profiles.",
+  "description": "An evaluation methodology for evaluating AI agent resilience under runtime tool-call disruptions, designed for the NeurIPS Evaluations & Datasets Track. Contains 100 base tasks and variants across 4 domains (Retail, Travel, Finance, DevOps) with 20 disruption types and 9 disruption profiles.",
   "conformsTo": "http://mlcommons.org/croissant/1.0",
   "url": "https://github.com/Kavirubc/AgentDisruptBench",
   "license": "https://opensource.org/licenses/MIT",