README.md

metadata

dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
  splits:
    - name: train_douban_sft
      num_bytes: 5567696
      num_examples: 3086
    - name: train_human_value_sft
      num_bytes: 806635
      num_examples: 1007
    - name: train_logi_qa_sft
      num_bytes: 666517
      num_examples: 421
    - name: train_ruozhiba_sft
      num_bytes: 228494
      num_examples: 240
    - name: train_segmentfault_sft
      num_bytes: 1068526
      num_examples: 458
    - name: train_wiki_sft
      num_bytes: 27611061
      num_examples: 10603
    - name: train_wikihow_sft
      num_bytes: 11069103
      num_examples: 1485
    - name: train_xhs_sft
      num_bytes: 2551884
      num_examples: 1508
    - name: train_zhihu_sft
      num_bytes: 13986060
      num_examples: 5631
    - name: train_sft
      num_bytes: 63555976
      num_examples: 24439
    - name: test_sft
      num_bytes: 228494
      num_examples: 240
  download_size: 67916523
  dataset_size: 127340446
configs:
  - config_name: default
    data_files:
      - split: train_douban_sft
        path: data/train_douban_sft-*
      - split: train_human_value_sft
        path: data/train_human_value_sft-*
      - split: train_logi_qa_sft
        path: data/train_logi_qa_sft-*
      - split: train_ruozhiba_sft
        path: data/train_ruozhiba_sft-*
      - split: train_segmentfault_sft
        path: data/train_segmentfault_sft-*
      - split: train_wiki_sft
        path: data/train_wiki_sft-*
      - split: train_wikihow_sft
        path: data/train_wikihow_sft-*
      - split: train_xhs_sft
        path: data/train_xhs_sft-*
      - split: train_zhihu_sft
        path: data/train_zhihu_sft-*
      - split: train_sft
        path: data/train_sft-*
      - split: test_sft
        path: data/test_sft-*
task_categories:
  - question-answering
language:
  - zh
pretty_name: e
size_categories:
  - 10K<n<100K

COIG-CQIA-sft: COIG-CQIA for sft in alignment-handbook

数据完全来自于COIG-CQIA。

暂时忽略了chinese_traditional，coig_pc，exam，finance这些转换麻烦或者语义上不适合当QA数据集的subset。 其中train_sft是全集，test_sft是ruozhiba，以便代码能够跑通。

经过reformat的本数据集可以直接使用alignment-handbook 进行sft。

可以使用zephyr-7b-beta/sft/config_qlora.yaml进行尝试，模型和数据集都可以使用本地文件夹。

@misc{bai2024coig,
    title={COIG-CQIA: Quality is All You Need for Chinese Instruction Fine-tuning}, 
    author={Bai, Yuelin and Du, Xinrun and Liang, Yiming and Jin, Yonggang and Liu, Ziqiang and Zhou, Junting and Zheng, Tianyu and Zhang, Xincheng and Ma, Nuo and Wang, Zekun and others},
    year={2024},
    eprint={2403.18058},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
  }