dingo-python 1.0

Language quality evaluation tool.

English | 简体中文

Introduction

Dingo is a data quality assessment tool that helps you automatically detect data quality issues in your datasets. Dingo provides a variety of built-in detection rules and model methods, and also supports custom detection methods. It supports commonly used NLP datasets and multimodal datasets, including pre-training datasets, fine-tuning datasets, and evaluation datasets. In addition, Dingo supports various interface usage methods, including local CLI, SDK, and RESTFul API, making it easy to integrate into various environments. The large model evaluation platform of OpenCompass has integrated Dingo for subjective evaluation of language model generated content.

Architecture of Dingo

QuickStart

Install dingo.

pip install dingo

Try the following demo code.

from dingo.model import Model
from dingo.io import RawInputModel
from dingo.exec import Executor

input_data = {
    "eval_models": ["default"],
    "input_path": "data/inputs/test_data1.json", # replace with your data path here
    "data_type": "json",
    "column_content": ["prediction"],
}

raw_input = RawInputModel(**input_data)
Model.apply_config(raw_input.custom_config_path)
executor = Executor.exec_map["local"](raw_input)
result = executor.evaluate()
print(result)

Tutorials

Config

Dingo enables users to personalize their data quality inspection methods, which can include the use of heuristic rules, third-party quality inspection tools or services, and large models. These can be implemented through configuration. Specifically, users can pass a parameter named custom_config_path which points to a configuration file. Below is a template for this configuration file: template.json

Rules

Heuristic rules are a common method for data processing and quality inspection. Dingo has implemented a series of heuristic rules and grouped them into rule groups, such as zh-all, and en-all, which represent the heuristic quality inspection rule sets for Chinese and English respectively. In the template of the configuration file, the two configuration items related to heuristic rule configuration are custom_rule_list and rule_config, which represent the rule set and the configuration parameters for a specific rule, respectively. Below is a configuration example:

{
  "custom_rule_list": [],
  "rule_config": {}
}

Large Models

Dingo supports data quality inspection using large models. Before use, users need to configure llm_config. For OpenAI models:

{
  "key": "YOUR_API_KEY"
}

For HuggingFace models(currently support the downloaded models):

{
  "path": "your local model path",
}

Execute

Dingo can be run locally or on a Spark cluster.

Local Mode

In addition to the aforementioned SDK calls, you can also run data evaluation locally with CLI:

python -m dingo.run.cli

The CLI parameters are as follows.

parameter name	description
-i or --input_path	The path of data. It can be a file or a directory.
-e or --eval_models	The model used to evaluate data quality.
--dataset_id	The id of data input.
--data_type	The type of data. It can be JSON, jsonl, plaintext and list json.
--output_path	The path of result data.
--column_id	The column name of id in data.
--column_prompt	The column name of prompt in data.
--column_content	The column name of content in data.
custom_config_path	The path of custom config file.

More information can be obtained by running the following command: python -m dingo.run.cli --help.

Spark Mode

If the scale of data is very large you can use Spark to run the project.

Firstly, create an object from SparkExecutor, and set the actual instances of SparkSession and DataFrame.

from dingo.exec.spark import SparkExecutor

spark_exec = SparkExecutor()
spark_exec.set_spark(spark_session)
spark_exec.set_input_df(spark_data_frame)

Then, convert the data and execute the rule list.

spark_exec.convert_data(column_id=['data_id'], column_prompt=['prompt'], column_content=['content'])
spark_exec.execute(["CommonSpecialCharacter", "CommonColonEnd"])

Finally, summarize and get the result data.

spark_exec.summarize()
output_df = spark_exec.get_output_df()
summary = spark_exec.get_summary()

Evaluation Results

Summary

The summary.json file is overall information about evaluation results. Here is an example:

{
    "dataset_id": "20240618",
    "input_model": "default",
    "input_path": "data/inputs/test_data1.json",
    "output_path": "data/outputs/20240625_134409",
    "score": 90.0,
    "num_good": 90,
    "num_bad": 10,
    "total": 100,
    "error_ratio": {...}
}

The error_ratio field shows data quality signals in seven different aspects: EFFECTIVENESS, COMPLETENESS, UNDERSTANDABILITY, SIMILARITY, FLUENCY, RELEVANCE and SECURITY.

Detailed Results

For more detailed issues found in data items, Dingo created files in a directory named with the quality signals mentioned above. Give an example. CommonColonEnd.json in the QUALITY_SIGNAL_COMPLETENESS directory is as follows:

{
    "name": "CommonColonEnd", # rule name
    "count": 1,
    "ratio": 0.5,
    "detail": [
        {
            "data_id": "0",
            "prompt": "",
            "content": "I am 8 years old. ^I love apple because:",
            "error_reason": "Ends with a colon."
        }
    ]
}

We evaluated the quality of these three datasets based on Dingo.

Dataset	Dataset Type	EFFECTIVENESS	COMPLETENESS	UNDERSTANDABILITY	SIMILARITY	FLUENCY	RELEVANCE	SECURITY
SlimPajama-627B	Pretrain	0	0.001797	0.011547	0.003563	0	0	0
Stanford_alpaca	SFT	0.0008	0.0004	0.0013	0.0002	0	0	0
MMLU	Benchmark	0.0064	0.0005	0.0113	0	0	0	0

Contributing

We appreciate all contributions to Dingo. Please refer to CONTRIBUTING.md for the contributing guideline.

License

This project is released under the Apache 2.0 license.