HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models

November 19, 2024 · View on GitHub

HelloBench is an open-source benchmark designed to evaluate the long text generation capabilities of large language models (LLMs). This repository includes the complete test data and evaluation code from the associated paper:

HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models

The test data are curated from platforms like Quora and Reddit, providing diverse, real-world challenges to evaluate LLM performance.

Merge to Opencompass

HelloBench is now merged to Opencompass, you can launch the test script automatically in Opencompass, more details can be found in url

Repository Contents

│  LICENSE
│  llm_judge.py
│  README.md
│  regression.py
│  requirements.txt
│  run.py
│
├─Checklists
│      chat_checklist.json
│      heuristic_text_generation_checklist.json
│      open_ended_qa_checklist.json
│      summarization_checklist.json
│      text_completion_checklist.json
│
├─Annotation_Interface
│      main.py
│      stats.jsonl
│
└─HelloBench
    │  chat.jsonl
    │  heuristic_text_generation.jsonl
    │  open_ended_qa.jsonl
    │  summarization.jsonl
    │  text_completion.jsonl
    │
    ├─length_constrained_experiments_data
	     heuristic_text_generation_16k.jsonl
       heuristic_text_generation_2k.jsonl
       heuristic_text_generation_4k.jsonl
       heuristic_text_generation_8k.jsonl

Setup Instructions

Language Requirements: Python 3.10 or later

To set up the environment, run the following command in your terminal:

pip install -r requirements.txt

Usage Guidelines

Test Data: The core test data can be found in the HelloBench directory.
Evaluation Checklists: For predefined evaluation criteria, refer to the Checklists directory.
Human Evaluation: To facilitate human evaluation, the required code is located in the Annotation_Interface directory.
LLM Response Generation: Use run.py to call an LLM to generate responses for the tasks.
LLM Response Judging: To invoke LLMs for evaluating responses, run llm_judge.py.
Regression: For Linear Regression, execute regression.py.

For additional details and advanced usage, please refer to the code comments and paper.

Generation Configuration

Here are the generation configurations for different models.

Model Name	max_new_tokens	temperature	version
GPT-4o-2024-0806	16384	0.8	gpt-4o-2024-08-06
Mistral-Large-API	16384	0.8	mistral-large-latest
o1-Mini	32768¹	0.8	o1-mini
Claude-3.5-Sonnet	8192²	0.8	claude-3-5-sonnet-20240620
Gemini-1.5-Pro	8192³	0.8	gemini-1.5-pro
Deepseek-API	4096⁴	0.8	deepseek-chat
Yi-Large	16384	0.8	yi-large
Qwen-Max	2000⁵	0.8	qwen-max-0428
GLM-4-API	4096⁶	0.8	glm-4-0520
Gemma-2-27B	4096⁷	0.8	google/gemma-2-27b-it
LLaMA-3.1-70B	16384	0.8	meta-llama/Meta-Llama-3.1-70B-Instruct
Qwen-2-72B	16384	0.8	Qwen/Qwen2-72B-Instruct
InternLM-2.5-20B	16384	0.8	internlm/internlm2_5-20b-chat
Yi-1.5-34B	2048⁸	0.8	01-ai/Yi-1.5-34B-Chat
LLaMA-3.1-8B	16384	0.8	meta-llama/Meta-Llama-3.1-8B-Instruct
GLM-4-9B	16384	0.8	THUDM/glm-4-9b-chat
Qwen-2-7B	16384	0.8	Qwen/Qwen2-7B-Instruct
InternLM-2.5-7B	16384	0.8	internlm/internlm2_5-7b-chat
Mistral-7B-0.2	16384	0.8	mistralai/Mistral-7B-Instruct-v0.2
Phi-3.5-Moe	16384	0.8	microsoft/Phi-3.5-MoE-instruct
MAP-Neo	2048⁹	0.8	m-a-p/neo_7b_instruct_v0.1
LongWriter-GLM4-9B	16384	0.8	THUDM/LongWriter-glm4-9b
Suri-I-ORPO	16384	0.8	chtmp223/suri-i-orpo
Yi-1.5-34B-16K	8192¹⁰	0.8	01-ai/Yi-1.5-34B-Chat-16K
InternLM-2.5-7B-1M	16384	0.8	internlm/internlm2_5-7b-chat-1m
GLM-4-9B-1M	16384	0.8	THUDM/glm-4-9b-chat-1m

For the o1-mini model, the parameter here should be max_completion_tokens instead of max_new_tokens, because it includes reasoning tokens. Therefore, I set it to 32768. You can refer to https://platform.openai.com/docs/guides/reasoning#controlling-costs for more details. ↩
For the reason that claude-3.5-sonnet has max output 8192 tokens. You can refer to https://docs.anthropic.com/en/docs/about-claude/models#model-comparison-table for more details. ↩
For the reason that gemini-1.5-pro has max output 8192 tokens. You can refer to https://ai.google.dev/gemini-api/docs/models/gemini#gemini-1.5-pro for more details. ↩
For the reason that deepseek-chat has max output 4096 tokens. You can refer to https://api-docs.deepseek.com/zh-cn/quick_start/pricing for more details. ↩
For the reason that qwen-max-0428 has max output 2000 tokens. You can refer to https://help.aliyun.com/zh/model-studio/getting-started/models?spm=a2c4g.11186623.0.0.74b04823IseC0N#9f8890ce29g5u for more details. ↩
For the reason that glm-4-0520 has max output 4096 tokens. You can refer to https://bigmodel.cn/dev/howuse/model for more details. ↩
Gemma-2-27B has max_position_embeddings with 8192 tokens, and you need to balance the input and output, thus result the selection of max_new_tokens=4096. ↩
Yi-1.5-34B-Chat has max_position_embeddings with 4096 tokens, and you need to balance the input and output, thus result the selection of max_new_tokens=2048. ↩
neo_7b_instruct_v0.1 has max_position_embeddings with 8192 tokens, and you need to balance the input and output, thus result the selection of max_new_tokens=2048. ↩
Yi-1.5-34B-Chat-16K has max_position_embeddings with 16384 tokens, and you need to balance the input and output, thus result the selection of max_new_tokens=8192. ↩

Merge to Opencompass

Repository Contents

Setup Instructions

Usage Guidelines

Generation Configuration

Footnotes