The Automatic Evaluation of LLMs (AEOLLM) task is a new core task in NTCIR-18 to support in-depth research on large language models (LLMs) evaluation. As LLMs grow popular in both fields of academia and industry, how to effectively evaluate the capacity of LLMs becomes an increasingly critical but still challenging issue. Existing methods can be divided into two types: manual evaluation, which is expensive, and automatic evaluation, which faces many limitations including the task format (the majority belong to multiple-choice questions) and evaluation criteria (occupied by reference-based metrics). To advance the innovation of automatic evaluation, we proposed the Automatic Evaluation of LLMs (AEOLLM) task which focuses on generative tasks and encourages reference-free methods. Besides, we set up diverse subtasks such as summary generation, non-factoid question answering, text expansion, and dialogue generation to comprehensively test different methods. We believe that the AEOLLM task will facilitate the development of the LLMs community.
""", unsafe_allow_html=True) elif page == "Methodology": st.header("Methodology") st.image("asserts/method.svg", use_column_width=True) st.markdown("""First, we choose four subtasks as shown in the table below:
| Task | Description | Dataset |
|---|---|---|
| Summary Generation (SG) | write a summary for the specified text | XSum: over 226k news articles |
| Non-Factoid QA (NFQA) | construct long-form answers to open-ended non-factoid questions | NF_CATS: 12k non-factoid questions |
| Text Expansion (TE) | given a theme, participants need to generate stories related to the theme | WritingPrompts: 303k story themes |
| Dialogue Generation (DG) | generate human-like responses to numerous topics in daily conversation contexts | DailyDialog: 13k daily conversation contexts |
Second, we choose a series of popular LLMs during the competition to generate answers.
Third, we manually annotate the answer sets for each question, which will be used as gold standards for evaluating the performance of different evaluation methods.
Last, we will collect evaluation results from participants and calculate consistency with manually annotated results. We will use Accuracy, Kendall’s tau and Spearman correlation coefficient as the evaluation metrics.
""",unsafe_allow_html=True) elif page == "Datasets": st.header("Answer Generation and Human Annotation") st.markdown(""" We randomly sampled **100 instances** from **each** dataset as the question set and selected **7 different LLMs** to generate answers, forming the answer set. As a result, each dataset produced 700 instances, totaling **2,800 instances across the four datasets**. For each instance (question-answer pair), we employed human annotators to provide a score ranging from 1 to 5 and took the median of these scores as the final score. Based on this score, we calculated the rankings of the 7 answers for each question. If scores were identical, the answers were assigned the same rank, with the lowest rank being used. """) st.header("Data Acquisition and Usage") st.markdown(""" We divided the 2,800 instances into three parts: - train set: 20% of the data (covering all four datasets) was designated as the training set (including human annotations) for participants to reference when designing their methods. - test set: Another 20% of the data was set aside as the test set (excluding human annotations), used to evaluate the performance of participants' methods and to generate the **leaderboard**. - reserved set: The remaining 60% of the data was reserved for **the final evaluation**. Both the training set and the test set can be downloaded from the provided link: [https://huggingface.co/datasets/THUIR/AEOLLM](https://huggingface.co/datasets/THUIR/AEOLLM). """) st.header("Resources") st.markdown("""A brief description of the specific dataset we used, along with the original download link, is provided below:
1. Summary Generation (SG): Xsum: A real-world single document news summary dataset collected from online articles by the British Broadcasting Corporation (BBC) and contains over 220 thousand news documents.
2. Non-Factoid QA (NFQA): NF_CATS: A dataset contains examples of 12k natural questions divided into eight categories.
3. Text Expansion (TE): WritingPrompts: A large dataset of 300K human-written stories paired with writing prompts from an online forum.
4. Dialogue Generation (DG): DailyDialog: A high-quality dataset of 13k multi-turn dialogues. The language is human-written and less noisy.
""",unsafe_allow_html=True) elif page == "Important Dates": st.header("Important Dates") st.markdown("""All deadlines are at 11:59pm in the Anywhere on Earth (AOE) timezone.
Kickoff Event: March 29, 2024During the Dry run (until Jan 15, 2025), we will use the test set to evaluate the performance of participants' methods and release the results on the Leaderboard.
Before the Formal run begins (before Jan 15, 2025), we will release the reserved set. Participants need to submit their results for the reserved set before the Formal run ends (before Feb 1, 2025).
We will be following a similar format as the ones used by most TREC submissions, which is repeated below. White space is used to separate columns. The width of the columns in the format is not important, but it is important to have exactly five columns per line with at least one space between the columns.
taskId questionId answerId score rank
1. the first column is the taskeId (index different tasks)
2. the second column is questionId (index different questions in the same task)
3. the third column is answerId (index the answer provided by different LLMs to the same question)
4. the fourth column is score (index the score to the answer given by participants)
5. the fifth column is rank (index the rank of the answer within all answers to the same question)
""",unsafe_allow_html=True) elif page == "Submit": st.header("Submit") st.markdown(""" Please organize the answers in a **txt** file, where each line includes: **taskId questionId answerId score rank**. Finally, name the file as **teamId_methods.txt** and submit it through the link below: [https://forms.gle/ctJD5tvZkYcCw7Kz9](https://forms.gle/ctJD5tvZkYcCw7Kz9) Each team can submit up to 5 times per day, and only the latest submission will be considered. The Leaderboard will be updated daily around 24:00 Beijing Time. A baseline example can be found in the [baseline_example](https://huggingface.co/spaces/THUIR/AEOLLM/tree/main/baseline_example) folder, where the output folder provides an [example](https://huggingface.co/spaces/THUIR/AEOLLM/blob/main/baseline_example/output/baseline1_chatglm3_6B.txt) of the submission file content. """) elif page == "LeaderBoard": st.header("LeaderBoard") # # 描述 st.markdown("""This leaderboard is used to show the performance of the automatic evaluation methods of LLMs submitted by the AEOLLM team on four tasks:
1. Dialogue Generation (DG)
2. Text Expansion (TE)
3. Summary Generation (SG)
4. Non-Factoid QA (NFQA)
The Leaderboard will be updated daily around 24:00 Beijing Time.
""", unsafe_allow_html=True) index = pd.MultiIndex.from_tuples([ ('', 'teamId'), ('', 'methods'), ('', 'overall'), ('Dialogue Generation', 'accuracy'), ('Dialogue Generation', "kendall's tau"), ('Dialogue Generation', 'spearman'), ('Text Expansion', "accuracy"), ('Text Expansion', "kendall's tau"), ('Text Expansion', 'spearman'), ('Summary Generation', 'accuracy'), ('Summary Generation', "kendall's tau"), ('Summary Generation', 'spearman'), ('Non-Factoid QA', "accuracy"), ('Non-Factoid QA', "kendall's tau"), ('Non-Factoid QA', 'spearman') ]) data = { ('', 'teamId'): ['baseline', 'baseline', 'baseline', 'baseline'], ('', 'methods'): ["chatglm3-6b", "baichuan2-13b", "chatglm-pro", "gpt-4o-mini"], ('', 'overall'): [0,0,0,0], ('Dialogue Generation', 'accuracy'): [0.5806, 0.5483, 0.6001, 0.6472], ('Dialogue Generation', "kendall's tau"): [0.3243, 0.1739, 0.3042, 0.4167], ('Dialogue Generation', 'spearman'): [0.3505, 0.1857, 0.3264, 0.4512], ('Text Expansion', "accuracy"): [0.5107, 0.5050, 0.5461, 0.5581], ('Text Expansion', "kendall's tau"): [0.1281, 0.0635, 0.2716, 0.3864], ('Text Expansion', 'spearman'): [0.1352, 0.0667, 0.2867, 0.4157], ('Summary Generation', 'accuracy'): [0.6504, 0.6014, 0.7162, 0.7441], ('Summary Generation', "kendall's tau"): [0.3957, 0.2688, 0.5092, 0.5001], ('Summary Generation', 'spearman'): [0.4188, 0.2817, 0.5403, 0.5405], ('Non-Factoid QA', "accuracy"): [0.5935, 0.5817, 0.7000, 0.7203], ('Non-Factoid QA', "kendall's tau"): [0.2332, 0.2389, 0.4440, 0.4235], ('Non-Factoid QA', 'spearman'): [0.2443, 0.2492, 0.4630, 0.4511] } # overall = [0, 0, 0, 0] # for d in data: # if d != ('', 'teamId') and d != ('', 'methods') and d != ('', 'overall'): # for i in range(4): # overall[i] += data[d][i] # overall = [i / (3*4) for i in overall] # data[('', 'overall')] = overall # for d in data: # if d != ('', 'teamId') and d != ('', 'methods'): # for col in range(len(data[d])): # data[d][col] = "{:.4f}".format(data[d][col]) # print(data) st.dataframe(data, use_container_width=True) # # teamId 唯一标识码 # DG = { # "teamId": ["baseline1", "baseline2", "baseline3", "baseline4"], # "methods": ["chatglm3-6b", "baichuan2-13b", "chatglm-pro", "gpt-4o-mini"], # "accuracy": [0.5806, 0.5483, 0.6001, 0.6472], # "kendall's tau": [0.3243, 0.1739, 0.3042, 0.4167], # "spearman": [0.3505, 0.1857, 0.3264, 0.4512] # } # df1 = pd.DataFrame(DG) # TE = { # "teamId": ["baseline1", "baseline2", "baseline3", "baseline4"], # "methods": ["chatglm3-6b", "baichuan2-13b", "chatglm-pro", "gpt-4o-mini"], # "accuracy": [0.5107, 0.5050, 0.5461, 0.5581], # "kendall's tau": [0.1281, 0.0635, 0.2716, 0.3864], # "spearman": [0.1352, 0.0667, 0.2867, 0.4157] # } # df2 = pd.DataFrame(TE) # SG = { # "teamId": ["baseline1", "baseline2", "baseline3", "baseline4"], # "methods": ["chatglm3-6b", "baichuan2-13b", "chatglm-pro", "gpt-4o-mini"], # "accuracy": [0.6504, 0.6014, 0.7162, 0.7441], # "kendall's tau": [0.3957, 0.2688, 0.5092, 0.5001], # "spearman": [0.4188, 0.2817, 0.5403, 0.5405], # } # df3 = pd.DataFrame(SG) # NFQA = { # "teamId": ["baseline1", "baseline2", "baseline3", "baseline4"], # "methods": ["chatglm3-6b", "baichuan2-13b", "chatglm-pro", "gpt-4o-mini"], # "accuracy": [0.5935, 0.5817, 0.7000, 0.7203], # "kendall's tau": [0.2332, 0.2389, 0.4440, 0.4235], # "spearman": [0.2443, 0.2492, 0.4630, 0.4511] # } # df4 = pd.DataFrame(NFQA) # df = [df1, df2, df3, df4] # for d in df: # for col in d.select_dtypes(include=['float64', 'int64']).columns: # d[col] = d[col].apply(lambda x: f"{x:.4f}") # 创建标签页 # tab1, tab2, tab3, tab4 = st.tabs(["DG", "TE", "SG", "NFQA"]) # with tab1: # st.markdown("""Task: Dialogue Generation; Dataset: DialyDialog
""", unsafe_allow_html=True) # st.dataframe(df1, use_container_width=True) # with tab2: # st.markdown("""Task: Text Expansion; Dataset: WritingPrompts
""", unsafe_allow_html=True) # st.dataframe(df2, use_container_width=True) # with tab3: # st.markdown("""Task: Summary Generation; Dataset: Xsum
""", unsafe_allow_html=True) # st.dataframe(df3, use_container_width=True) # with tab4: # st.markdown("""Task: Non-Factoid QA; Dataset: NF_CATS
""", unsafe_allow_html=True) # st.dataframe(df4, use_container_width=True) st.markdown("A baseline example can be found in the [baseline_example](https://huggingface.co/spaces/THUIR/AEOLLM/tree/main/baseline_example) folder.") # 获取北京时间 time_placeholder = st.empty() beijing_tz = pytz.timezone('Asia/Shanghai') beijing_time = datetime.now(beijing_tz) while True: # 获取当前的北京时间 beijing_time = datetime.now(beijing_tz) # 在页面上动态显示当前北京时间 time_placeholder.write("Current Beijing Time: " + beijing_time.strftime('%Y-%m-%d %H:%M:%S')) # 设置更新频率为每秒钟一次 time.sleep(1) elif page == "Organisers": st.header("Organisers") st.markdown("""
Yiqun Liu [yiqunliu@tsinghua.edu.cn] (Tsinghua University)
Qingyao Ai [aiqy@tsinghua.edu.cn] (Tsinghua University)
Junjie Chen [chenjj826@gmail.com] (Tsinghua University)
Zhumin Chu [chuzm19@mails.tsinghua.edu.cn] (Tsinghua University)
Haitao Li [liht22@mails.tsinghua.edu.cn] (Tsinghua University)
Please feel free to contact us! 😉
[1] Mao R, Chen G, Zhang X, et al. GPTEval: A survey on assessments of ChatGPT and GPT-4. pdf
[2] Chang Y, Wang X, Wang J, et al. A survey on evaluation of large language models. pdf
[3] Chan C M, Chen W, Su Y, et al. Chateval: Towards better llm-based evaluators through multi-agent debate. pdf
[4] Li R, Patel T, Du X. Prd: Peer rank and discussion improve large language model based evaluations. pdf
[5] Chu Z, Ai Q, Tu Y, et al. Pre: A peer review based large language model evaluator. pdf