README_en.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - question-answering
tags:
  - agent
  - benchmark
  - tool-use
  - korean

🇰🇷 Ko-AgentBench v1

"Korean Agent Benchmark Project"

English | 한국어

As AI agents become more sophisticated, it has become crucial to precisely measure their performance under conditions similar to real-world environments. However, most benchmarks are designed based on English-speaking environments, which limits their ability to reflect Korea's unique usage contexts.

To address this issue, we have developed a high-quality agent benchmark specialized for the Korean real-world usage environment.

Ko-AgentBench Key Features ✨

1. Step-by-step Task Design

We have comprehensively analyzed agent capabilities across 7 levels, from simple tool calls to long-term contextual abilities and robustness handling capabilities.

2. 18 Korean-specific APIs and High-quality Scenarios Tailored to Real-life Environments

Based on APIs from Korean real-world usage environments such as Naver, Maps, Kakao, and websites, we have implemented realistic problem-solving scenarios closely related to domestic users' daily lives, such as 'appointment booking' and 'blog review search'.

3. Cache-based Iterative Evaluation and Robustness Testing

We solve chronic problems of existing benchmarks, such as 'information attribute inconsistency changes'. By improving failed API responses, we ensure benchmark consistency and reliability.

By evaluating error recognition/response capabilities (strategies) in intentional error situations, we select models that operate stably even in real-world environments.

4. Step-specific Precision Metrics

We evaluate the necessity/requirements of problem-solving step by step, including tool selection, parameter configuration, and data flow. Through this, we quantitatively identify the strengths and weaknesses of models.

Data Loading

from datasets import load_dataset

# Load specific level
dataset = load_dataset("Hugging-Face-KREW/Ko-AgentBench", data_files="L1.json")

# Or load all levels
dataset = load_dataset("Hugging-Face-KREW/Ko-AgentBench", data_files="*.json")

Dataset Overview

• Define task classification system for agent benchmark design
• Design to evaluate agent's tool calling capabilities in a step-by-step manner

Dataset Scope

• Evaluation Target: Open-weight sLLM (supports tool calling), Commercial APIs
• Evaluation Scope: Agent tool calling performance in single-turn and multi-turn conversation situations
• Applied APIs: 18 Korean-specific open APIs

Task Levels

Single-Turn

L1. Single Tool Call

• Goal: Verify the most basic API calling capability
• Description: Check if the given tool can be executed with correct parameters
• Feature: Evaluate "accuracy only" by performing requests with specified API names or natural language requests as-is
• Example: "Search for 'Rapid Current' using Naver Book API and tell me the price."
• Example: "Tell me the price of the 'Rapid Current' book"

L2. Tool Selection

• Goal: Verify the ability to select the optimal API among multiple candidate tools
• Description: Users make requests in natural language, and the model must select the most suitable tool from the given tool list
• Feature: Evaluate accurate tool mapping with input natural language
• Example: "Check the price of the 'All Back English Middle 2-1 Cheonjae (Kim)' book."
• Candidate tools: hotel_booking_api, aladin_books_api
  • Candidate tools must have no mutual correlation.

L3. Sequential Tool Reasoning

• Goal: Verify planning and execution capabilities through multi-step reasoning
• Description: Check if a correct pipeline can be constructed by connecting the results of one tool as input to another tool
• Feature: Evaluate "planned chain-of-tools" rather than simple calls
• Example: "Tell me when the Instax11 I bought from 11st Amazon will be delivered"
• Candidate tools: 11st_order_api, customs_api, cj_delivery_api
  • Tools must be callable sequentially (11st delivery number inquiry → customs clearance → courier company)

L4. Parallel Tool Reasoning

• Goal: Collect information in parallel and derive conclusions by synthesizing it
• Description: Simultaneously call multiple independent tools, compare and analyze results, then produce final answers
• Feature: Evaluate multi-source aggregation (information synthesis and comparison ability)
• Example: "Check the stock of the 'Hanroro Grapefruit Apricot Club' book."
• Candidate tools: kyobo_books_api, aladin_books_api
• Expected answer: There are 12 books at Kyobo Book Centre and 18 books at Aladin, totaling 30 books.
  • At this time, candidate tools must handle the same function in parallel.

L5. Error Handling and Robustness

• Goal: Verify coping ability in error situations
• Description: Evaluate how various failure modes are handled, not just "failed"
• Sub-items:
• A. Request for additional questions
  • Guide users to make clearer requests when information is insufficient
• B. Hallucination prevention
  • Prohibit calling non-existent APIs
  • Prohibit "pretending to succeed" answers when failed
• C. Fallback maneuvers
  • Whether alternative APIs with the same function can be utilized when specific API errors occur
• Example: "When Naver Movie API call fails → Report 'API call failed' or call Kakao Movie API as alternative"

Multi-Turn

L6. Efficient Tool Utilization

• Goal: Verify the ability to efficiently reuse previous tool results
• Description: While recalling APIs in all situations is accurate, it's inefficient in terms of cost and delay. Conversely, unconditionally reusing old information also causes accuracy problems.
• Feature: Evaluate whether reasonable choices can be made between "recall vs reuse"
• Example:
  • User: "Compare Coupang and Naver prices." → Result: Coupang 80, Naver 85
  • User: "What was the Naver price?"
  • Correct answer: 85 (utilize past information, avoid unnecessary recalls)
  • Wrong answer: Call API again or "I don't know"

L7. Long-Context Reasoning

• Goal: Verify the ability to maintain long-term context in multi-turn conversations
• Description: Remember information from several turns ago and correctly perform tool calling by connecting it with new questions
• Example:
  • User's first question: "I'm going to travel to Jeju Island."
  • Later: "How's the weather?" → Call weather API using Jeju Island context
  • (Additional turn) "If it rains, find places where I can buy an umbrella." → Utilize all previous Jeju Island + weather context

Links

You can check more detailed information about Ko-AgentBench.

• 🏆 Live Leaderboard
• 📊 Dataset
• 📝 Github

Contact

If you have any questions about the dataset and benchmark, please contact us!

Hugging Face KREW is a Korean non-profit research organization that strives to deeply understand artificial intelligence through Hugging Face and contribute to open source.

• ✍🏻 Blog: KREW-blog
• 🐦 HuggingFace Community: @huggingface-KREW
• 💼 LinkedIn: Hugging Face KREW