huangyt/FINETUNE1

📔 DATASET

Dataset	Class	Number of Questions
FLAN_CoT(zs)	Reasoning 、 MATH 、 ScienceQA 、 Commonsense	91910
Prm800k	Reasoning 、 MATH	6713
ScienceQA	ScienceQA	5177
SciBench	ScienceQA	695
ReClor	Reasoning	1624
TheoremQA	Commonsense 、 MATH 、 ScienceQA	800
OpenBookQA	Text_Understanding 、 Reasoning 、 Commonsense 、 ScienceQA	5957
ARB	Reasoning 、 MATH 、 ScienceQA 、 Commonsense 、 Text_Understanding	605
Openassistant-guanaco	Commonsense 、 Text_Understanding 、 Reasoning	802
SQuAD 2.0	Text_Understanding	87599
CommonsenseQA	Commonsense	9741
Ethics	Commonsense	21759

📌 Methon

Dataset Format Definition

Use "instruction、input、output" tend to lean towards guided datasets. In this format, each sample includes an instruction, an input, and an expected output. The instruction provides guidance on how to process the input to generate the output. This format of dataset is often used to train models to perform specific tasks, as they explicitly indicate the operations the model should perform.

  {
      "input": "",
      "output": "",
      "instruction": ""
  }

• FLAN_V2 COT(ZS)

  We only extract the 'zs_opt' from COT and categorize each task.

• CommonsenseQA

  We extracted the question and choices from the original CommonsenseQA dataset and placed them in the instruction. We also wrote the input prompt: "Choose A, B, C, D, or E as your solution."

• SQuAD

  We used the questions from the SQUAD dataset as instructions and treated the context as the input.

• Ethics

  The ethics dataset, which was originally in labeled format, has been transformed into a true or false format. Additionally, the input now includes the instruction "Give true or false according to ethics."

• OTHER

  Prm800k, ScienceQA, SciBench, ReClor, TheoremQA, OpenBookQA, ARB, and OpenAssistant-Guanaco datasets adopt the same format as Platypus.

Sampling Algorithms

1. First,we are taking all datasets from COT, ARB, TheoremQA and Ethics. ARB and TheoremQA encompass a wide range of fields and have a relatively low total count. Since COT has high quality, we are including the entire dataset. For the Ethics dataset, we are collecting the entire dataset because we want the model to comprehensively learn more about ethics and security aspects.

2. The remaining datasets were initially categorized into the following four groups for the purpose of Simple Random Sampling:

   • Science Questions and Answers : ScienceQA、SciBench
   • Reasoning & Mathematics : ReClor、Prm800k
   • Text Comprehension : OpenBookQA、SQuAD
   • Commonsense : CommonsenseQA、Openassistant-guanaco

   However, we discovered that the total number of datasets in the Science Questions and Answers、Reasoning & Mathematics、and Commonsense categories did not exceed 30,000. As a result, only the Text Comprehension category underwent Simple Random Sampling, while the others were taken in their entirety.

🏁 Feature Work

• In the future, we intend to utilize Stratified Sampling due to the imbalance in the number of questions across different datasets, which introduces bias. Conversely, if we opt to randomly sample an equal number of examples from each dataset, it can yield a smaller estimation error for the same total sample size.
• We can even evaluate based on the fine-tuning from the first stage and employ additional scripting techniques to enhance the quality of the dataset.