Nejumi Leaderboard 3

Overview

This repository is for the Nejumi Leaderboard 3, a comprehensive evaluation platform for large language models. The leaderboard assesses both general language capabilities and alignment aspects. For detailed information about the leaderboard, please visit Nejumi Leaderboard website.

Evaluation Metrics

Our evaluation framework incorporates a diverse set of metrics to provide a holistic assessment of model performance:

Main Category	Subcategory	Automated Evaluation with Correct Data	AI Evaluation	Note
General Language Processing	Expression		MT-bench/roleplay (0shot) MT-bench/humanities (0shot) MT-bench/writing (0shot)
^	Translation	ALT e-to-j (jaster) (0shot, 2shot) ALT j-to-e (jaster) (0shot, 2shot) wikicorpus-e-to-j(jaster) (0shot, 2shot) wikicorpus-j-to-e(jaster) (0shot, 2shot)
^	Summarization
^	Information Extraction	JSQuaD (jaster) (0shot, 2shot)
^	Reasoning		MT-bench/reasoning (0shot)
^	Mathematical Reasoning	MAWPS*(jaster) (0shot, 2shot) MGSM*(jaster) (0shot, 2shot)	MT-bench/math (0shot)
^	(Entity) Extraction	wiki_ner*(jaster) (0shot, 2shot) wiki_coreference(jaster) (0shot, 2shot) chABSA*(jaster) (0shot, 2shot)	MT-bench/extraction (0shot)
^	Knowledge / Question Answering	JCommonsenseQA*(jaster) (0shot, 2shot) JEMHopQA*(jaster) (0shot, 2shot) JMMLU*(0shot, 2shot) NIILC*(jaster) (0shot, 2shot) aio*(jaster) (0shot, 2shot)	MT-bench/stem (0shot)
^	English	MMLU_en (0shot, 2shot)
^	semantic analysis	JNLI*(jaster) (0shot, 2shot) JaNLI*(jaster) (0shot, 2shot) JSeM*(jaster) (0shot, 2shot) JSICK*(jaster) (0shot, 2shot) Jamp*(jaster) (0shot, 2shot)
^	syntactic analysis	JCoLA-in-domain*(jaster) (0shot, 2shot) JCoLA-out-of-domain*(jaster) (0shot, 2shot) JBLiMP*(jaster) (0shot, 2shot) wiki_reading*(jaster) (0shot, 2shot) wiki_pas*(jaster) (0shot, 2shot) wiki_dependency*(jaster) (0shot, 2shot)
Alignment	Controllability	jaster* (0shot, 2shot) LCTG		LCTG cannot be used for business purposes. Usage for research and using the result in the press release are acceptable.
^	Ethics/Moral	JCommonsenseMorality*(2shot)
^	Toxicity		LINE Yahoo Reliability Evaluation Benchmark	This dataset is not publicly available due to its sensitive content.
^	Bias	JBBQ (2shot)		JBBQ needs to be downloaded from JBBQ github repository.
^	Truthfulness	JTruthfulQA		For JTruthfulQA evaluation, nlp-waseda/roberta_jtruthfulqa requires Juman++ to be installed beforehand. You can install it by running the script/install_jumanpp.sh script.
^	Robustness	Test multiple patterns against JMMLU (W&B original) (0shot, 2shot) - Standard method - Choices are symbols - Select anything but the correct answer

• metrics with (0, 2-shot) are averaged across both settings.
• Metrics marked with an asterisk (*) evaluate control capabilities.
• For MT-bench, StabilityAI's MT-Bench JP is used with GPT-4o-2024-05-13 as the model to evaluate.
• For LCTG, the only quantity test is conducted. (The quality test is not conducted)
• vLLM is leveraged for efficient inference.
• Alignment data may contain sensitive information and the default setting does not include it in this repository. If you want to evaluate your models agains Alinghment data, please check each dataset instruction carefully

Implementation Guide

Environment Setup

1. Set up environment variables

export WANDB_API_KEY=<your WANDB_API_KEY>
export OPENAI_API_KEY=<your OPENAI_API_KEY>
export LANG=ja_JP.UTF-8
# If using Azure OpenAI instead of standard OpenAI
export AZURE_OPENAI_ENDPOINT=<your AZURE_OPENAI_ENDPOINT>
export AZURE_OPENAI_API_KEY=<your AZURE_OPENAI_API_KEY>
export OPENAI_API_TYPE=azure
# if needed, set the following API KEY too
export ANTHROPIC_API_KEY=<your ANTHROPIC_API_KEY>
export GOOGLE_API_KEY=<your GOOGLE_API_KEY>
export COHERE_API_KEY=<your COHERE_API_KEY>
export MISTRAL_API_KEY=<your MISTRAL_API_KEY>
export AWS_ACCESS_KEY_ID=<your AWS_ACCESS_KEY_ID>
export AWS_SECRET_ACCESS_KEY=<your AWS_SECRET_ACCESS_KEY>
export AWS_DEFAULT_REGION=<your AWS_DEFAULT_REGION>
export UPSTAGE_API_KEY=<your UPSTAGE_API_KEY>
# if needed, please login in huggingface
huggingface-cli login

2. Clone the repository

git clone https://github.com/wandb/llm-leaderboard.git
cd llm-leaderboard

3. Set up a Python environment with requirements.txt

Dataset Preparation

For detailed instructions on dataset preparation and caveate, please refer to scripts/data_uploader/README.md.

In Nejumi Leadeboard3, the following dataset are used.

Please ensure to thoroughly review the terms of use for each dataset before using them.

1. jaster(Apache-2.0 license)
2. MT-Bench-JA (Apache-2.0 license)
3. LCTG (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Permission from AI shift to use for the leaderboard and was received.)
4. JBBQ (Creative Commons Attribution 4.0 International License.)
5. LINE Yahoo Inappropriate Speech Evaluation Dataset (not publically available)
6. JTruthfulQA (Creative Commons Attribution 4.0 International License.)

Configuration

Base configuration

The base_config.yaml file contains basic settings, and you can create a separate YAML file for model-specific settings. This allows for easy customization of settings for each model while maintaining a consistent base configuration.

Below, you will find a detailed description of the variables utilized in the base_config.yaml file.

• wandb: Information used for Weights & Biases (W&B) support.

  • entity: Name of the W&B Entity.
  • project: Name of the W&B Project.
  • run_name: Name of the W&B run. Please set up run name in a model-specific config.

• testmode: Default is false. Set to true for lightweight implementation with a small number of questions per category (for functionality checks).

• inference_interval: Set inference interval in seconds. This is particularly effective when there are rate limits, such as with APIs.

• run: Set to true for each evaluation dataset you want to run.

• model: Information about the model.

  • artifacts_path: Path of the wandb artifacts where the model is located.
  • max_model_len: Maximum token length of the input.
  • chat_template: Path to the chat template file. This is required for open-weights models.
  • dtype: Data type. Choose from float32, float16, bfloat16.
  • trust_remote_code: Default is true.
  • device_map: Device map. Default is "auto".
  • load_in_8bit: 8-bit quantization. Default is false.
  • load_in_4bit: 4-bit quantization. Default is false.

• generator: Settings for generation. For more details, refer to the generation_utils in Hugging Face Transformers.

  • top_p: top-p sampling. Default is 1.0.
  • temperature: The temperature for sampling. Default is 0.1.
  • max_tokens: Maximum number of tokens to generate. This value will be overwritten in the script.

• num_few_shots: Number of few-shot examples to use.

• github_version: For recording, not required to be changed.

• jaster: Settings for the Jaster dataset.

  • artifacts_path: URL of the WandB Artifact for the Jaster dataset.
  • dataset_dir: Directory of the Jaster dataset after downloading the Artifact.

• jmmlu_robustness: Whether to include the JMMLU Robustness evaluation. Default is True.

• lctg: Settings for the LCTG dataset.

  • artifacts_path: URL of the WandB Artifact for the LCTG dataset.
  • dataset_dir: Directory of the LCTG dataset after downloading the Artifact.

• jbbq: Settings for the JBBQ dataset.

  • artifacts_path: URL of the WandB Artifact for the JBBQ dataset.
  • dataset_dir: Directory of the JBBQ dataset after downloading the Artifact.

• toxicity: Settings for the toxicity evaluation.

  • artifact_path: URL of the WandB Artifact of the toxicity dataset.
  • judge_prompts_path: URL of the WandB Artifact of the toxicity judge prompts.
  • max_workers: Number of workers for parallel processing.
  • judge_model: Model used for toxicity judgment. Default is gpt-4o-2024-05-13

• jtruthfulqa: Settings for the LCTG dataset.

  • artifact_path: URL of the WandB Artifact for the JTruthfulQA dataset.
  • roberta_model_name: Name of the RoBERTa model used for evaluation. Default is 'nlp-waseda/roberta_jtruthfulqa'.

• mtbench: Settings for the MT-Bench evaluation.

  • temperature_override: Override the temperature for each category of the MT-Bench.
  • question_artifacts_path: URL of the WandB Artifact for the MT-Bench questions.
  • referenceanswer_artifacts_path: URL of the WandB Artifact for the MT-Bench reference answers.
  • judge_prompt_artifacts_path: URL of the WandB Artifact for the MT-Bench judge prompts.
  • bench_name: Choose 'japanese_mt_bench' for the Japanese MT-Bench, or 'mt_bench' for the English version.
  • model_id: The name of the model. You can replace this with a different value if needed.
  • question_begin: Starting position for the question in the generated text.
  • question_end: Ending position for the question in the generated text.
  • max_new_token: Maximum number of new tokens to generate.
  • num_choices: Number of choices to generate.
  • num_gpus_per_model: Number of GPUs to use per model.
  • num_gpus_total: Total number of GPUs to use.
  • max_gpu_memory: Maximum GPU memory to use (leave as null to use the default).
  • dtype: Data type. Choose from None, float32, float16, bfloat16.
  • judge_model: Model used for judging the generated responses. Default is gpt-4o-2024-05-13
  • mode: Mode of evaluation. Default is 'single'.
  • baseline_model: Model used for comparison. Leave as null for default behavior.
  • parallel: Number of parallel threads to use.
  • first_n: Number of generated responses to use for comparison. Leave as null for default behavior.

Model configuration

After setting up the base-configuration file, the next step is to set up a configuration file for model under configs/.

API Model Configurations

This framework supports evaluating models using APIs such as OpenAI, Anthropic, Google, and Cohere. You need to create a separate config file for each API model. For example, the config file for OpenAI's gpt-4o-2024-05-13 would be named configs/config-gpt-4o-2024-05-13.yaml.

• wandb: Information used for Weights & Biases (W&B) support.
  • run_name: Name of the W&B run.
• api: Choose the API to use from openai, anthropic, google, amazon_bedrock.
• batch_size: Batch size for API calls (recommended: 32).
• model: Information about the model.
  • pretrained_model_name_or_path: Name of the API model.
  • size_category: Specify "api" to indicate using an API model.
  • size: Model size (leave as null for API models).
  • release_date: Model release date. (MM/DD/YYYY)

Other Model Configurations

This framework also supports evaluating models using VLLM. You need to create a separate config file for each VLLM model. For example, the config file for Microsoft's Phi-3-medium-128k-instruct would be named configs/config-Phi-3-medium-128k-instruct.yaml.

• wandb: Information used for Weights & Biases (W&B) support.
  • run_name: Name of the W&B run.
• api: Set to vllm to indicate using a VLLM model.
• num_gpus: Number of GPUs to use.
• batch_size: Batch size for VLLM (recommended: 256).
• model: Information about the model.
  • artifacts_path: When loading a model from wandb artifacts, it is necessary to include a description. If not, there is no need to write it. Example notation: wandb-japan/llm-leaderboard/llm-jp-13b-instruct-lora-jaster-v1.0:v0
  • pretrained_model_name_or_path: Name of the VLLM model.
  • chat_template: Path to the chat template file (if needed).
  • size_category: Specify model size category. In Nejumi Leaderboard, the category is defined as "10B<", "10B<= <30B", "<=30B" and "api".
  • size: Model size (parameter).
  • release_date: Model release date (MM/DD/YYYY).
  • max_model_len: Maximum token length of the input (if needed).

Create Chat template (needed for models except for API)

1. create chat_templates/model_id.jinja If the chat_template is specified in the tokenizer_config.json of the evaluation model, create a .jinja file with that configuration. If chat_template is not specified in tokenizer_config.json, refer to the model card or other relevant documentation to create a chat_template and document it in a .jinja file.

2. test chat_templates If you want to check the output of the chat_templates, you can use the following script:

python3 scripts/test_chat_template.py -m <model_id> -c <chat_template>

If the model ID and chat_template are the same, you can omit -c <chat_template>.

Evaluation Execution

Once you prepare the dataset and the configuration files, you can run the evaluation process.

You can use either -c or -s option: - -c (config): Specify the config file by its name, e.g., python3 scripts/run_eval.py -c config-gpt-4o-2024-05-13.yaml - -s (select-config): Select from a list of available config files. This option is useful if you have multiple config files.

python3 scripts/run_eval.py -s
or 
python3 scripts/run_eval.py -c

The results of the evaluation will be logged to the specified W&B project.

When you want to edit runs or add additional evaluation metrics

Please refer to belend_run_configs/README.md.

Contributing

Contributions to this repository is welcom. Please submit your suggestions via pull requests. Please note that we may not accept all pull requests.

License

This repository is available for commercial use. However, please adhere to the respective rights and licenses of each evaluation dataset used.

Contact

For questions or support, please concatct to contact-jp@wandb.com.