hakkero-dataloader 1.2.23

hakkero-dataloader

A general dataloader build on top of Pytorch Dataloader.

1. How to use

1.1 Build Index

Install pip install hakkero-dataloader and run the following command to build index.

hakkero -h

usage: hakkero [-h] [--version] [--filename FILENAME] [--output OUTPUT] --dtype {legacy,message,preference} [--num_workers NUM_WORKERS] [--not_shuf]

build index for dataset

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit
  --filename FILENAME   full filename of jsonl file
  --output OUTPUT       output path for saving data.jsonl and index.h5
  --dtype {legacy,message,preference}
                        data type
  --num_workers NUM_WORKERS
                        number of workers
  --not_shuf            not shuf data

1.2 Use In Training

from hakkero.dataset import get_data

dp_world_size, dp_rank = 1, 0
tokenizer = ...
batch_size = 4
max_length = 4096
n_workers = 2

dataset, dataloader, forward_keys = get_data(
    config="/path/to/dataset config",
    dp_rank=dp_rank,
    dp_world_size=dp_world_size,
    tokenizer=tokenizer,
    batch_size=batch_size,
    max_length=max_length,
    # segment and tokenize strategy or set them in `config` and let strategy_segment=None and strategy_tokenize=None: 
    st_segment="naive",
    st_tokenize="legacy",
    # add bos/eos token for legacy tokenize strategy
    add_bos_token=True,
    add_eos_token=True,
    # norm dataset weight with tokens of target
    norm_weight_with_n_targets=False,
    # keep <think>xxx</think> in message or not
    # no - not keep <think>xx</think> (default)
    # last - keep <think>xx</think> in last turn
    # all - keep <think>xx</think> in all turns
    keep_think="no",
    homogeneous=True,
    seed=9527,
    n_workers=n_workers,
    is_preference=False,
    use_unpad_data=False,
    use_unpad_in_pad=False,
    packed=False,  # 样本是否已经 packed  过了
    shift=True  # tokenize 返回是否 shift 1 了（input[:-1], label[1:])
)

prefetcher = dataloader.prefetch(n_workers, drop_last=False)
for step, batch in enumerate(prefetcher, start=0):
  print(batch)

example of config:

{
    "hermes25_1":
    {
        "group": "en",
        "name": "hermes25_1",
        "epoch": 1,
        "path": "hermes25",
        "strategy":
        {
            "st_segment": "integrous",
            "st_tokenize": "hg"
        },
        "weight": 0.5
    },
    "hermes25_2":
    {
        "group": "en",
        "name": "hermes25_1",
        "epoch": 1,
        "path": "hermes25",
        "strategy":
        {
            "st_segment": "integrous",
            "st_tokenize": "hg"
        },
        "weight": 0.5
    }
}

2. Supported Strategies

See segmentation.py and tokenization.py for more details.

2.1 Segmentation Strategies

• integrous: discard sample that is too long, exceed max_length
• concat: split long sample, concat it with previous segment, shuffle all segments
  • not support preference data.
• naive: split long sample with random length, shuffle all segments
  • not support preference data.
• unbiased: split long sample exceed max_length with random length, shuffle all segments.
  • not support preference data.

2.2 Tokenization Strategies

• legacy: \n\n as delimiter to join text and use tokenizer.encode to encode the input.

  • format of input data

    {
      "uid": "xxx",
      "data":
      {
          "title": "xxx",
          "summary": "xxx",
          "abstract": "xxx",
          "text": "xxx",
          "question": "xxx",
          "answer": "xxx",
          "code": "xxx",
          "label": "xxx"
      }
    }
    

    • All fields except label are stripped and joined with "\n\n" as the context.
    • label is the target to learn for finetuning (pretrain data should not have the label field).
    • See func legacy in tokenization.py for more details.

  • extra parameters:

    • add_bos_token, add_eos_token

• hg: huggingface message data, use tokenizer.apply_chat_template to encode the input.

  • format of input data

    {
      "uid": "xx",
      "data": [
        {"role": "user", "content": "xxx"},
        {"role": "assistant", "content": "xxx"},
         ...
      ]
    }
    

    See func huggingface_message in tokenization.py for more details.

  • extra parameters:

    • keep_think: support keep <think>xx</think> or not
      • no - not keep <think>xx</think> (default)
      • last - keep <think>xx</think> in last turn
      • all - keep <think>xx</think> in all turns

• chatml: chat message data, use chatml to encode the input.

  • format of input data

    {
      "uid": "xx",
      "data": [
        {"role": "user", "content": "xxx"},
        {"role": "assistant", "content": "xxx"},
         ...
      ]
    }
    

    See func chatml_message in tokenization.py for more details.

• chatml_qwen2_vl_message: chat message vl data, use chatml to encode the input.

  • format of input data

    {
      "uid": "xx",
      "data": [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": "images/2.jpg"
                },
                {
                    "type": "text",
                    "text": "他是谁？"
                }
            ]
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "text",
                    "text": "他是来自拜仁慕尼黑的托马斯·穆勒。"
                }
            ]
        },
         ...
      ]
    }
    

    See func chatml_qwen2_vl_message in tokenization.py for more details. Only support "integrous" segmentation strategies

• hg_preference: preference data, use tokenizer.apply_chat_template to encode the input.

  • format of input data

    {
      "uid": "xx",
      "data": {
        "context": [
          {"role": "user", "content": "xxx"},
          {"role": "assistant", "content": "xxx"},
          ...
          {"role": "user", "content": "xxx"}
        ],
        "chosen": "chosen response",
        "rejected": "rejected response"
      }
    }
    

    See func huggingface_preference in tokenization.py for more details.

  • extra parameters:

    • keep_think: support keep <think>xx</think> or not
      • no - not keep <think>xx</think> (default)
      • last - keep <think>xx</think> in last turn
      • all - keep <think>xx</think> in all turns

• chatml_preference: preference data, use chatml to encode the input.

  • format of input data

    {
      "uid": "xx",
      "data": {
        "context": [
          {"role": "user", "content": "xxx"},
          {"role": "assistant", "content": "xxx"},
          ...
          {"role": "user", "content": "xxx"}
        ],
        "chosen": "chosen response",
        "rejected": "rejected response"
      }
    }
    

    See func chatml_preference in tokenization.py for more details.