CRAFT
收藏资源简介:
CRAFT是一个用于任务特定合成数据集生成的项目,通过语料库检索和增强技术,提供了多个领域的合成数据集,如生物医学问答、常识问答、医学问答等。
CRAFT is a project focused on task-specific synthetic dataset generation. Leveraging corpus retrieval and augmentation techniques, it provides synthetic datasets across multiple domains, including biomedical question answering, commonsense question answering, medical question answering, and so on.
CRAFT: Corpus Retrieval and Augmentation for Fine-Tuning
数据集概述
CRAFT项目提供了多个合成数据集,这些数据集适用于不同的任务,包括生物学问答(BioQA)、常识问答(CommonSenseQA, CSQA)、医学问答(MedQA)、食谱生成(RecipeGen)和摘要生成(Summarization)。这些数据集在Hugging Face平台上可用。
数据集链接
- BioQA: https://huggingface.co/datasets/ingoziegler/CRAFT-BioQA
- CommonSenseQA (CSQA): https://huggingface.co/datasets/ingoziegler/CRAFT-CommonSenseQA
- MedQA: https://huggingface.co/datasets/ingoziegler/CRAFT-MedQA
- RecipeGen: https://huggingface.co/datasets/ingoziegler/CRAFT-RecipeGen
- Summarization: https://huggingface.co/datasets/ingoziegler/CRAFT-Summarization
数据集使用
要使用人工编写的小样本数据,可以通过过滤数据集中的is_few_shot == 1来获取,或者直接加载assets/{task}/few-shot/corpus-task-32.jsonl文件。8个小样本实验使用每个文件的前8行。
性能表现
在合成数据集上训练的模型能够匹配通用指令调优的大型语言模型(LLMs)的性能,甚至在某些任务(如摘要生成)上超过人类策划数据的训练效果。合成数据在面对分布偏移时更为稳健,因为数据不是为特定测试集生成,而是为整体任务生成。
5-gram重叠率比较
| BioQA | CSQA | MedQA | Summarization | |
|---|---|---|---|---|
| CRAFT<sub>XS</sub> | 0.0% | 0.0% | 0.0% | 0.0% |
| CRAFT<sub>S</sub> | 0.0% | 0.1% | 0.1% | 0.0% |
| CRAFT<sub>M</sub> | 0.0% | 0.2% | 0.1% | 0.1% |
| CRAFT<sub>L</sub> | 0.0% | 0.4% | 0.3% | 0.2% |
| CRAFT<sub>XL</sub> | 0.0% | 0.2% | 0.2% | 0.2% |
| Baseline <small>(In-domain Train Set)</small> | 17.9% | 4.4% | 1.1% | 0.3% |
跨测试集性能比较
| Dataset | Baseline | CRAFT<sub>XL</sub> |
|---|---|---|
| In-domain | 89.9 | 78.1 |
| MMLU<sub>Medical Genetics</sub> | 60.0 | 69.0 |
| MMLU<sub>Anatomy</sub> | 55.6 | 57.0 |
| MMLU<sub>High School Biology</sub> | 69.3 | 67.4 |
| MMLU<sub>College Biology</sub> | 66.7 | 74.3 |
| MMLU-Avg | 62.9 | 66.9 |
适配器检查点
提供了在CRAFT-XL版本上微调后得到的适配器检查点的下载链接:
- BioQA: https://huggingface.co/ingoziegler/CRAFT-BioQA-XL
- CommonSenseQA (CSQA): https://huggingface.co/ingoziegler/CRAFT-CommonSenseQA-XL
- MedQA: https://huggingface.co/ingoziegler/CRAFT-MedQA-XL
- RecipeGen: https://huggingface.co/ingoziegler/CRAFT-RecipeGen-XL
- Summarization: https://huggingface.co/ingoziegler/CRAFT-Summarization-XL




