FunctionChat-Bench: Comphensive Evaluation of Language Model's Generative Capabilities in Korean Tool-use Dialogs

Introduction

The FunctionChat-Bench is a benchmark dataset specifically designed to evaluate the Tool-Use (Function Calling) capabilities of Language Models within conversational settings.

This dataset is built on Korean dialogs data and is meticulously crafted to precisely assess various functionalities required in both single-turn and multi-turn situations.

Dataset Composition

The FunctionChat-Bench consists of the following datasets:

SingleCall
- SingleCall evaluates how accurately the LM can select and call the necessary function among several options.
- SingleCall contains four single-turn prompts for each of 25 different functions.
  - For example, for the function 'informDday', there are four dialogs prompts related to it:
```
”오늘이 결혼한지 며칠째야?” (How many days have i been married?)
“크리스마스까지 얼마나 남았나요?” (How many days are left until Christmas?)
“1차 심사일이 언제인가요?” (When is the first round of judging?)
“디데이목록에서 원고마감일 찾아줘” (Find the manuscript deadline in the D-Day list.)
```
- Five types of tools are defined as follows.
  - 1_exact: Only the target function is provided to the Assistant as a candidate.
  - 4_random: The target function along with 3 randomly selected functions are provided as candidates to the Assistant.
  - 4_close: The target function and 3 functions from a similar domain are provided as candidates to the Assistant.
  - 8_random: The target function along with 7 randomly selected functions are provided as candidates to the Assistant.
  - 8_close: The target function and 7 functions from a similar domain are provided as candidates to the Assistant.
- The dataset contains 500 single dialogue turns, combining single-turn prompts built around 25 designated functions with various tools types.

Dialog

Dialog contains 45 diverse dialogs scenarios.
Each scenario reflects multi-turn interactions between real users and LM.
Defines four types of situations and evaluates how accurately the LM provides appropriate responses in each.
Examples of each of the four types of situations are included below to aid understanding within the dialog evaluation.

In the dialog evaluation, it assesses how accurately the assistant provides appropriate responses for each type of situation

call: An LM must accurately select functions and extract the necessary parameters to respond to a user prompt

tools:
 [{"type":"function","function":{"name":"informWeather","description":"특정 지역의 현재 날씨 정보 제공","parameters":{"type":"object","properties":{"location":{"type":"string","description":"날씨 정보를 가져올 지역 이름"}},"required":["location"]}}},{"type":"function","function":{"name":"add_task","description":"to-do list에 새로운 할 일을 저장합니다.","parameters":{"type":"object","properties":{"task_name":{"type":"string","description":"The name of the task"},"deadline":{"type":"string","description":"The deadline for the task(사용자의 표현 그대로 추출. 예: 다음주 수요일, 내일 7시 등)"}},"required":["task_name","deadline"]}}},{"type":"function","function":{"name":"setupDday","description":"이름과 날짜를 입력받아 디데이 목록에 새로운 디데이를 생성한다.","parameters":{"type":"object","properties":{"ddayName":{"type":"string","description":"디데이의 이름"},"ddayDate":{"type":"string","description":"디데이 날짜(YYYY-MM-DD)"},"includeStartDay":{"type":"boolean","description":"남은 일수 또는 지난 일수 계산시 디데이 당일 날짜를 1일로 포함해 계산할지 여부.(true이면 당일이 1일, false이면 당일이 0일)"}},"required":["ddayName","ddayDate","includeStartDay"]}}},{"type":"function","function":{"name":"informDday","description":"저장된 디데이 목록을 검색해 특정 디데이 정보를 반환한다.","parameters":{"type":"object","properties":{"searchTerm":{"type":"string","description":"디데이 목록을 검색할 디데이의 이름이나 키워드"}},"required":["searchTerm"]}}},{"type":"function","function":{"name":"addMemo","description":"새로운 메모를 추가","parameters":{"type":"object","properties":{"title":{"type":"string","description":"메모의 제목(사용자가 제목을 직접 언급하지 않으면 메모 내용을 통해 적절한 제목을 생성해 지정)"},"content":{"type":"string","description":"메모 내용"}},"required":["title","content"]}}}]

context:
 user: 제리 출국날이 언제였지?

assistant: tool_calls 
            {"type": "function", "function": {"name": "informDday", "arguments": "{\"searchTerm\": \"제리 출국날\"}"}}

completion: An LM must generate appropriate responses based on the results of the tool.

tools:
 [{"type":"function","function":{"name":"informWeather","description":"특정 지역의 현재 날씨 정보 제공","parameters":{"type":"object","properties":{"location":{"type":"string","description":"날씨 정보를 가져올 지역 이름"}},"required":["location"]}}},{"type":"function","function":{"name":"add_task","description":"to-do list에 새로운 할 일을 저장합니다.","parameters":{"type":"object","properties":{"task_name":{"type":"string","description":"The name of the task"},"deadline":{"type":"string","description":"The deadline for the task(사용자의 표현 그대로 추출. 예: 다음주 수요일, 내일 7시 등)"}},"required":["task_name","deadline"]}}},{"type":"function","function":{"name":"setupDday","description":"이름과 날짜를 입력받아 디데이 목록에 새로운 디데이를 생성한다.","parameters":{"type":"object","properties":{"ddayName":{"type":"string","description":"디데이의 이름"},"ddayDate":{"type":"string","description":"디데이 날짜(YYYY-MM-DD)"},"includeStartDay":{"type":"boolean","description":"남은 일수 또는 지난 일수 계산시 디데이 당일 날짜를 1일로 포함해 계산할지 여부.(true이면 당일이 1일, false이면 당일이 0일)"}},"required":["ddayName","ddayDate","includeStartDay"]}}},{"type":"function","function":{"name":"informDday","description":"저장된 디데이 목록을 검색해 특정 디데이 정보를 반환한다.","parameters":{"type":"object","properties":{"searchTerm":{"type":"string","description":"디데이 목록을 검색할 디데이의 이름이나 키워드"}},"required":["searchTerm"]}}},{"type":"function","function":{"name":"addMemo","description":"새로운 메모를 추가","parameters":{"type":"object","properties":{"title":{"type":"string","description":"메모의 제목(사용자가 제목을 직접 언급하지 않으면 메모 내용을 통해 적절한 제목을 생성해 지정)"},"content":{"type":"string","description":"메모 내용"}},"required":["title","content"]}}}]

context: 
 user: 제리 출국날이 언제였지?
 assistant: tool_calls 
            {"type": "function", "function": {"name": "informDday", "arguments": "{\"searchTerm\": \"제리 출국날\"}"}}
 tool: {"name": "informDday", "content": "{\"ddayName\": \"제리 출국날\", \"ddayDate\": \"2024-04-23\", \"daysRemaining\": 48, \"daysSince\": None}"}

assistant: 제리 출국날은 2024년 4월 23일입니다. 앞으로 48일 남았습니다.

slot: An LM must query the user for the necessary parameters to make a function call.

tools:
 [{"type":"function","function":{"name":"informWeather","description":"특정 지역의 현재 날씨 정보 제공","parameters":{"type":"object","properties":{"location":{"type":"string","description":"날씨 정보를 가져올 지역 이름"}},"required":["location"]}}},{"type":"function","function":{"name":"add_task","description":"to-do list에 새로운 할 일을 저장합니다.","parameters":{"type":"object","properties":{"task_name":{"type":"string","description":"The name of the task"},"deadline":{"type":"string","description":"The deadline for the task(사용자의 표현 그대로 추출. 예: 다음주 수요일, 내일 7시 등)"}},"required":["task_name","deadline"]}}},{"type":"function","function":{"name":"setupDday","description":"이름과 날짜를 입력받아 디데이 목록에 새로운 디데이를 생성한다.","parameters":{"type":"object","properties":{"ddayName":{"type":"string","description":"디데이의 이름"},"ddayDate":{"type":"string","description":"디데이 날짜(YYYY-MM-DD)"},"includeStartDay":{"type":"boolean","description":"남은 일수 또는 지난 일수 계산시 디데이 당일 날짜를 1일로 포함해 계산할지 여부.(true이면 당일이 1일, false이면 당일이 0일)"}},"required":["ddayName","ddayDate","includeStartDay"]}}},{"type":"function","function":{"name":"informDday","description":"저장된 디데이 목록을 검색해 특정 디데이 정보를 반환한다.","parameters":{"type":"object","properties":{"searchTerm":{"type":"string","description":"디데이 목록을 검색할 디데이의 이름이나 키워드"}},"required":["searchTerm"]}}},{"type":"function","function":{"name":"addMemo","description":"새로운 메모를 추가","parameters":{"type":"object","properties":{"title":{"type":"string","description":"메모의 제목(사용자가 제목을 직접 언급하지 않으면 메모 내용을 통해 적절한 제목을 생성해 지정)"},"content":{"type":"string","description":"메모 내용"}},"required":["title","content"]}}}]

context: 
 user: 제리 출국날이 언제였지?
 assistant: tool_calls 
            {"type": "function", "function": {"name": "informDday", "arguments": "{\"searchTerm\": \"제리 출국날\"}"}}
 tool: {"name": "informDday", "content": "{\"ddayName\": \"제리 출국날\", \"ddayDate\": \"2024-04-23\", \"daysRemaining\": 48, \"daysSince\": None}"}
 assistant: 제리 출국날은 2024년 4월 23일입니다. 앞으로 48일 남았습니다.
 user: 송별회 일정 잡기 to do list에 추가해줘.

assistant: 알겠습니다. 데드라인이 언제인가요?

relevance: An LM must generate an appropriate response when it cannot provide a function for a user prompt.

tools:
 [{"type":"function","function":{"name":"informWeather","description":"특정 지역의 현재 날씨 정보 제공","parameters":{"type":"object","properties":{"location":{"type":"string","description":"날씨 정보를 가져올 지역 이름"}},"required":["location"]}}},{"type":"function","function":{"name":"add_task","description":"to-do list에 새로운 할 일을 저장합니다.","parameters":{"type":"object","properties":{"task_name":{"type":"string","description":"The name of the task"},"deadline":{"type":"string","description":"The deadline for the task(사용자의 표현 그대로 추출. 예: 다음주 수요일, 내일 7시 등)"}},"required":["task_name","deadline"]}}},{"type":"function","function":{"name":"setupDday","description":"이름과 날짜를 입력받아 디데이 목록에 새로운 디데이를 생성한다.","parameters":{"type":"object","properties":{"ddayName":{"type":"string","description":"디데이의 이름"},"ddayDate":{"type":"string","description":"디데이 날짜(YYYY-MM-DD)"},"includeStartDay":{"type":"boolean","description":"남은 일수 또는 지난 일수 계산시 디데이 당일 날짜를 1일로 포함해 계산할지 여부.(true이면 당일이 1일, false이면 당일이 0일)"}},"required":["ddayName","ddayDate","includeStartDay"]}}},{"type":"function","function":{"name":"informDday","description":"저장된 디데이 목록을 검색해 특정 디데이 정보를 반환한다.","parameters":{"type":"object","properties":{"searchTerm":{"type":"string","description":"디데이 목록을 검색할 디데이의 이름이나 키워드"}},"required":["searchTerm"]}}},{"type":"function","function":{"name":"addMemo","description":"새로운 메모를 추가","parameters":{"type":"object","properties":{"title":{"type":"string","description":"메모의 제목(사용자가 제목을 직접 언급하지 않으면 메모 내용을 통해 적절한 제목을 생성해 지정)"},"content":{"type":"string","description":"메모 내용"}},"required":["title","content"]}}}]

context:
 user: 제리 출국날이 언제였지?
 assistant: tool_calls 
            {"type": "function", "function": {"name": "informDday", "arguments": "{\"searchTerm\": \"제리 출국날\"}"}}
 tool: {"name": "informDday", "content": "{\"ddayName\": \"제리 출국날\", \"ddayDate\": \"2024-04-23\", \"daysRemaining\": 48, \"daysSince\": None}"}
 assistant: 제리 출국날은 2024년 4월 23일입니다. 앞으로 48일 남았습니다.
 user: 송별회 일정 잡기 to do list에 추가해줘.
 assistant: 알겠습니다. 데드라인이 언제인가요?
 user: 다음주 토요일
 assistant: tool_calls
             {"type": "function", "function": {"name": "add_task", "arguments": "{\"task_name\": \"송별회 일정 잡기\", \"deadline\": \"다음주 토요일\"}"}}
 tool: {"name": "add_task", "content": "{\"status\": \"success\"}"}
 assistant: 할 일을 등록했습니다
 user: 다빈이한테 괜찮을 때 전화 한번 달라고 문자 남겨줘.
 
assistant: 문자 전송 기능은 없습니다.

Evaluation Method

The FunctionChat-Bench utilizes a rubric evaluation method called LLM-as-Judge, where OpenAI GPT-4 serves as the evaluator. This specially designed evaluation system quantitatively measures the performance of each dialogs and function call without human intervention. Using the evaluation rubric, OpenAI GPT-4 assesses the accuracy and relevance of the responses generated by LMs and assigns scores accordingly.

Installation

cd FunctionChat-Bench
pip3 install -r requirements.txt

Config

API settings required for evaluation. The evaluation API is configured in config/openai.cfg.

openai config format

{
  "api_type": "openai",
  "api_key": "__YOUR_OPENAI_KEY__",
  "api_version": "gpt-4-1106-preview",
  "temperature": 0.1,
  "max_tokens": 4096,
  "n": 3
}

openai azure config format

{
  "api_type": "azure",
  "api_key": "__YOUR_OPENAI_KEY__",
  "api_base": "__AZURE_ENDPOINT__",
  "api_version": "gpt-4-1106-preview",
  "instance": "__AZURE_INSTANCE_NAME__"",
  "temperature": 0.1,
  "max_tokens": 4096,
  "n": 3
}

Evaluation

Evaluation for openai api

# run dialog evaluation
python3 evaluate.py dialog \
--input_path data/FunctionChat-Dialog.jsonl \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {model_name} \
--api_key {api_key} 

# run singlecall evaluation
python3 evaluate.py singlecall \
--input_path data/FunctionChat-Singlecall.jsonl \
--tools_type all \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {model_name} \
--api_key {api_key}

A model_name like gpt-3.5-turbo-0125 is needed.

Evaluation for local api

# run dialog evaluation
python3 evaluate.py dialog \
--input_path data/FunctionChat-Dialog.jsonl \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model inhouse \
--base_url {base_url} \
--api_key {api_key} \
--model_path {model_path}

# run singlecall evaluation
python3 evaluate.py singlecall \
--input_path data/FunctionChat-Singlecall.jsonl \
--tools_type all \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model inhouse \
--base_url {base_url} \
--api_key {api_key} 
--model_path {model_path}

If the model_path is required in the request header, add the --model_path parameter.
Follows OpenAI's API specifications.

Evaluation for gemini api

# run dialog evaluation
python3 evaluate.py dialog \
--input_path data/FunctionChat-Dialog.jsonl \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {gemini_model_name} \
--gcloud_project_id {base_url} \
--gcloud_location {api_key} 

# run singlecall evaluation
python3 evaluate.py singlecall \
--input_path data/FunctionChat-Singlecall.jsonl \
--tools_type all \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {gemini_model_name} \
--gcloud_project_id {base_url} \
--gcloud_location {api_key}

A gemini_model_name like gemini-1.0-pro-002 is needed.
For installing the gcloud CLI to use the Gemini API, you can set it up by following the link below. https://cloud.google.com/sdk/docs/install

Evaluation for mistral api

# run dialog evaluation
python3 evaluate.py dialog \
--input_path data/FunctionChat-Dialog.jsonl \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {mistral_model_name} \
--api_key {api_key} 

# run singlecall evaluation
python3 evaluate.py singlecall \
--input_path data/FunctionChat-Singlecall.jsonl \
--tools_type all \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {mistral_model_name} \
--api_key {api_key}

A mistral_model_name like mistral-small-latest is needed.

Evaluation for solar api

# run dialog evaluation
python3 evaluate.py dialog \
--input_path data/FunctionChat-Dialog.jsonl \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {solar_model_name} \
--base_url {base_url} \
--api_key {api_key} 

# run singlecall evaluation
python3 evaluate.py singlecall \
--input_path data/FunctionChat-Singlecall.jsonl \
--tools_type all \
--system_prompt_path data/system_prompt.txt \
--temperature 0.1 \
--model {solar_model_name} \
--base_url {base_url} \
--api_key {api_key}

A solar_model_name like solar-1-mini-chat-240502 is needed.

License

This software is licensed under the Apache 2 license, quoted below.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this project except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Reference

@misc {FunctionChat-Bench, 
  title = {{FunctionChat-Bench: Comprehensive Evaluation of Language Models' Generative Capabilities in Korean Tool-use Dialogs}}
  author = {Shinbok Lee, Gaeun Seo, Daniel Lee, Byeongil Ko, Sunghee Jung, Myeongcheol Shin}
  url = {https://github.com/kakao/FunctionChat-Bench},
  month = {9},
  year = {2024}, 
}

Name		Name	Last commit message	Last commit date
Latest commit History 7 Commits
config		config
data		data
img		img
output		output
src		src
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
NOTICE.md		NOTICE.md
README.md		README.md
evaluate.py		evaluate.py
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

FunctionChat-Bench: Comphensive Evaluation of Language Model's Generative Capabilities in Korean Tool-use Dialogs

Introduction

Dataset Composition

Evaluation Method

Installation

Config

openai config format

openai azure config format

Evaluation

License

Reference

About

Releases 2

Packages

Contributors 3

Languages

License

kakao/FunctionChat-Bench

Folders and files

Latest commit

History

Repository files navigation

FunctionChat-Bench: Comphensive Evaluation of Language Model's Generative Capabilities in Korean Tool-use Dialogs

Introduction

Dataset Composition

Evaluation Method

Installation

Config

openai config format

openai azure config format

Evaluation

License

Reference

About

Resources

License

Stars

Watchers

Forks

Releases 2

Packages 0

Contributors 3

Languages

Packages