Skip to main content

A flexible data transformation tool for ML training formats (SFT, RLHF, Pretrain)

Project description

dtflow

简洁的数据格式转换工具,专为机器学习训练数据设计。

安装

pip install dtflow

# 可选依赖
pip install tiktoken          # Token 统计(OpenAI 模型)
pip install transformers      # Token 统计(HuggingFace 模型)
pip install datasets          # HuggingFace Dataset 转换

快速开始

from dtflow import DataTransformer

# 加载数据
dt = DataTransformer.load("data.jsonl")

# 链式操作:过滤 -> 转换 -> 保存
(dt.filter(lambda x: x.score > 0.8)
   .to(lambda x: {"q": x.question, "a": x.answer})
   .save("output.jsonl"))

核心功能

数据加载与保存

# 支持 JSONL、JSON、CSV、Parquet
dt = DataTransformer.load("data.jsonl")
dt.save("output.jsonl")

# 从列表创建
dt = DataTransformer([{"q": "问题", "a": "答案"}])

数据过滤

# Lambda 过滤
dt.filter(lambda x: x.score > 0.8)

# 支持属性访问
dt.filter(lambda x: x.language == "zh")

数据转换

# 自定义转换
dt.to(lambda x: {"question": x.q, "answer": x.a})

# 使用预设模板
dt.to(preset="openai_chat", user_field="q", assistant_field="a")

预设模板

预设名称 输出格式
openai_chat {"messages": [{"role": "user", ...}, {"role": "assistant", ...}]}
alpaca {"instruction": ..., "input": ..., "output": ...}
sharegpt {"conversations": [{"from": "human", ...}, {"from": "gpt", ...}]}
dpo_pair {"prompt": ..., "chosen": ..., "rejected": ...}
simple_qa {"question": ..., "answer": ...}

Token 统计

from dtflow import count_tokens, token_counter, token_filter, token_stats

# 计算 token 数量
count = count_tokens("Hello world", model="gpt-4")

# 添加 token_count 字段
dt.transform(token_counter("text")).save("with_tokens.jsonl")

# 按 token 长度过滤
dt.filter(token_filter("text", max_tokens=2048))
dt.filter(token_filter(["question", "answer"], min_tokens=10, max_tokens=4096))

# 统计 token 分布
stats = token_stats(dt.data, "text")
# {"total_tokens": 12345, "avg_tokens": 123, "min_tokens": 5, "max_tokens": 500, ...}

支持 tiktoken(OpenAI,默认)和 transformers 后端:

# 使用 transformers tokenizer
count_tokens("Hello", model="Qwen/Qwen2-7B", backend="transformers")

格式转换器

from dtflow import (
    to_hf_dataset, from_hf_dataset,    # HuggingFace Dataset
    to_openai_batch, from_openai_batch, # OpenAI Batch API
    to_llama_factory,                   # LLaMA-Factory 格式
    to_axolotl,                         # Axolotl 格式
    messages_to_text,                   # messages 转纯文本
)

# HuggingFace Dataset 互转
ds = to_hf_dataset(dt.data)
ds.push_to_hub("my-dataset")

data = from_hf_dataset("tatsu-lab/alpaca", split="train")

# OpenAI Batch API
batch_input = dt.to(to_openai_batch(model="gpt-4o"))
results = from_openai_batch(batch_output)

# 训练框架格式
dt.transform(to_llama_factory()).save("llama_factory.jsonl")
dt.transform(to_axolotl()).save("axolotl.jsonl")

# messages 转纯文本(支持 chatml/llama2/simple 模板)
dt.transform(messages_to_text(template="chatml"))

其他操作

# 采样
dt.sample(100)           # 随机采样 100 条
dt.head(10)              # 前 10 条
dt.tail(10)              # 后 10 条

# 分割
train, test = dt.split(ratio=0.8, shuffle=True, seed=42)

# 统计
stats = dt.stats()       # 总数、字段信息
count = dt.count(lambda x: x.score > 0.9)

# 打乱
dt.shuffle(seed=42)

CLI 命令

# 数据采样
dt sample data.jsonl --num=10
dt sample data.csv --num=100 --sample_type=head

# 数据转换 - 预设模式
dt transform data.jsonl --preset=openai_chat
dt transform data.jsonl --preset=alpaca

# 数据转换 - 配置文件模式
dt transform data.jsonl                    # 首次运行生成配置文件
# 编辑 .dt/data.py 后再次运行
dt transform data.jsonl --num=100          # 执行转换

# 数据清洗
dt clean data.jsonl --drop-empty                    # 删除任意空值记录
dt clean data.jsonl --drop-empty=text,answer        # 删除指定字段为空的记录
dt clean data.jsonl --min-len=text:10               # text 字段最少 10 字符
dt clean data.jsonl --max-len=text:1000             # text 字段最多 1000 字符
dt clean data.jsonl --keep=question,answer          # 只保留这些字段
dt clean data.jsonl --drop=metadata                 # 删除指定字段
dt clean data.jsonl --strip                         # 去除字符串首尾空白
dt clean data.jsonl --strip --drop-empty=text --min-len=text:10 -o clean.jsonl  # 组合使用

# 数据去重
dt dedupe data.jsonl                            # 全量精确去重
dt dedupe data.jsonl --key=text                 # 按字段精确去重
dt dedupe data.jsonl --key=text --similar=0.8   # 相似度去重

# 文件拼接
dt concat a.jsonl b.jsonl -o merged.jsonl

# 数据统计
dt stats data.jsonl

错误处理

# 跳过错误项(默认)
dt.to(transform_func, on_error="skip")

# 抛出异常
dt.to(transform_func, on_error="raise")

# 保留原始数据
dt.to(transform_func, on_error="keep")

# 返回错误信息
result, errors = dt.to(transform_func, return_errors=True)

设计哲学

函数式优于类继承

不需要复杂的 OOP 抽象,直接用函数解决问题:

# ✅ 简单直接
dt.to(lambda x: {"q": x.question, "a": x.answer})

# ❌ 不需要这种设计
class MyFormatter(BaseFormatter):
    def format(self, item): ...

预设是便利层,不是核心抽象

90% 的需求用 transform(lambda x: ...) 就能解决。预设只是常见场景的快捷方式:

# 预设:常见场景的便利函数
dt.to(preset="openai_chat")

# 自定义:完全控制转换逻辑
dt.to(lambda x: {
    "messages": [
        {"role": "user", "content": x.q},
        {"role": "assistant", "content": x.a}
    ]
})

KISS 原则

  • 一个核心类 DataTransformer 搞定所有操作
  • 链式 API,代码像自然语言
  • 属性访问 x.field 代替 x["field"]
  • 不过度设计,不追求"可扩展框架"

实用主义

不追求学术上的完美抽象,只提供足够好用的工具

License

MIT

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dtflow-0.2.0.tar.gz (51.3 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

dtflow-0.2.0-py3-none-any.whl (50.1 kB view details)

Uploaded Python 3

File details

Details for the file dtflow-0.2.0.tar.gz.

File metadata

  • Download URL: dtflow-0.2.0.tar.gz
  • Upload date:
  • Size: 51.3 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for dtflow-0.2.0.tar.gz
Algorithm Hash digest
SHA256 a7d915fc5046fdc116e35ddba112c07953e77bb6fef52cef3d51a46a30eab370
MD5 dd2eddb6888ed3353184b1fa2bfb75e4
BLAKE2b-256 33cba135c09553951185e06346ca0afa2f2e2ea713ebafa3e66a694f3a111745

See more details on using hashes here.

Provenance

The following attestation bundles were made for dtflow-0.2.0.tar.gz:

Publisher: python-publish.yml on ai-ivan/data_transformer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file dtflow-0.2.0-py3-none-any.whl.

File metadata

  • Download URL: dtflow-0.2.0-py3-none-any.whl
  • Upload date:
  • Size: 50.1 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.7

File hashes

Hashes for dtflow-0.2.0-py3-none-any.whl
Algorithm Hash digest
SHA256 1af234d4d4e191b9882ed2573406b6cf25acc788a75ad1f835f64d0b4e75ec8c
MD5 36216c8e0c20cfac7774cc387a248ba7
BLAKE2b-256 35f88015a948bf8b149ebc67c7866133478aeed302b074105d6bdde84a96234f

See more details on using hashes here.

Provenance

The following attestation bundles were made for dtflow-0.2.0-py3-none-any.whl:

Publisher: python-publish.yml on ai-ivan/data_transformer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page