LCTG Bench: LLM Controlled Text Generation Benchmark

The LCTG Bench has been built to measure the controllability of Japanese LLMs in terms of how well they comply with constraints such as character count keywords in instructions.

LCTG Bench has been constructed in collaboration with AI Shift Inc., CyberAgent Inc. and Okazaki Lab at Tokyo Institute of Technology.

Task/Datasets

The LCTG Bench consists of three text generation tasks: Summarization, Advertisement (AD) Text Geneartion, and Pros & Cons Generation. For each task, the performance of LLM is evaluated in terms of four controllability aspects: Format, Character Count, Keyword, and Prohibited word.

All data sets consist of test data only.

The Pros & Cons Generation task has been built after writing the Japanese paper.

Task	Dataset	Format	Character Count	Keyword	Prohibihited word
Summarization	ABEMA Times	120	120	120	120
Ad Text Generation	CAMERA	150	150	150	150
Pros & Cons Generation	-	150	150	150	150

Evaluation Method

You can run the LLM performance evaluation using this benchmark simply by executing run_lctg.sh. All scripts are placed under scripts, and datasets are placed under datasets.

Procedure

0. To run this script, you will need to prepare a python 3.10 operating environment in advance. (We have not verified that it works with python 3.11 or later versions.)
1. To install packages, run the following command.

pip install -r requirements.txt

2. To conduct an evaluation, you must prepare a .env file with OPENAI_API_KEY, VERSION, and MODEL_PATH_X (where X= 0, 1, 2 ...) defined (please specify "1" for VERSION). For MODEL_PATH_X, specify the path of the model on the hugging face hub or the name of the model provided by the API. Please always specify X as a sequential number starting from 0.

OPENAI_API_KEY={your_openai_api_key}
VERSION="1"
MODEL_PATH_0="gpt-4-1106-preview"
MODEL_PATH_1="cyberagent/calm2-7b-chat"

3. By executing the following command, inference by LLM and evaluation of the inference results will be conducted. After execution, you can check the scores at output/{task}/all_scores.csv. It takes several hours to execute the evaluation for each model.

make run

Notes

Currently, this repository covers the following models.

MODEL_PATH_0="gpt-4-1106-preview"
MODEL_PATH_1="gpt-3.5-turbo-0125"
MODEL_PATH_2="gemini-pro"
MODEL_PATH_3="cyberagent/calm2-7b-chat"
MODEL_PATH_4="elyza/ELYZA-japanese-Llama-2-7b-fast-instruct"
MODEL_PATH_5="line-corporation/japanese-large-lm-3.6b-instruction-sft"
MODEL_PATH_6="matsuo-lab/weblab-10b-instruction-sft"
MODEL_PATH_7="rinna/youri-7b-chat"
MODEL_PATH_8="stabilityai/japanese-stablelm-instruct-gamma-7b"
MODEL_PATH_9="llm-jp/llm-jp-13b-instruct-full-jaster-v1.0"
MODEL_PATH_10="llm-jp/llm-jp-13b-instruct-full-jaster-dolly-oasst-v1.0"
MODEL_PATH_11="tokyotech-llm/Swallow-7b-instruct-hf"
MODEL_PATH_12="tokyotech-llm/Swallow-13b-instruct-hf"
MODEL_PATH_13="tokyotech-llm/Swallow-70b-instruct-hf"

If you wish to evaluate the performance of models other than those listed above, you will need to edit the following files in addition to adding environment variables. But if you want to use the openai model. You only have to edit scripts/utils.model_info_getter.py

scripts/utils/tokenizer_model_tokenids.py

• you have to add a function that returns model and tokenizer

def get_calm2_model_tokenizer(model_path:str):
    model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", torch_dtype="auto")
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    return model, tokenizer

scripts/utils/output_tokens.py

• You have to add a function that returns the output tokens (str).
• You may also have to define system prompts and model parameters in the functions you add.

def get_calm2_output_tokens(model, tokenizer, text: str, use_system_prompt: bool = True):
    prompt = f"""USER: {text}
ASSISTANT: """
    token_ids = tokenizer.encode(prompt, return_tensors="pt")
    with torch.no_grad():
        output_ids = model.generate(
            input_ids=token_ids.to(model.device),
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=True,
            temperature=0.8,
        )
        output_ids = output_ids[0][len(token_ids[0]):]
    return tokenizer.decode(output_ids, skip_special_tokens=True)

scripts/utils.model_info_getter.py

• You have to edit the following two functions.
  • get_model_tokenizer
    • you have to add few lines.

    elif "cyberagent/calm2" in model_path:
        model, tokenizer = get_calm2_model_tokenizer(model_path)

• get_output_tokens
  • You have to add the pair of key (model name) and value (function name)

"cyberagent/calm2-7b-chat": get_calm2_output_tokens,

Citation

If you use it in your research, please cite:

@InProceedings{Kurihara_nlp2024,
  author = 	"栗原健太郎 and 三田雅人 and 張培楠 and 佐々木翔大 and 石上亮介 and 岡崎直観",
  title = 	"LCTG Bench: 日本語LLMの制御性ベンチマークの構築",
  booktitle = 	"言語処理学会第30回年次大会",
  year =	"2024",
  url = "https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/D11-2.pdf"
  note= "in Japanese"
}

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Prohibited Uses

The evaluation results from this benchmark may be made available to the public, but the following uses are prohibited

• Mentioning Abema TV Inc. as the provider of ABEMA TIMES, the data source for the summarization task