Data Preparation Bench

A unified, downstream-grounded benchmark for data preparation in LLM training pipelines.

---

👋 Overview

DataPrep-Bench is the first unified, downstream-grounded benchmark that jointly evaluates how well LLMs, agents, and data workflows can prepare training data end to end. It covers three complementary tracks:

• Data Construction — transforms raw sources into SFT data.
• Data Selection — selects high-utility subsets from candidate pools.
• Data Quality Evaluation — predicts the downstream utility of candidate datasets.

It also ships strong baselines for each track: Data-Construction-Skill for skill-driven agentic construction, a unified Data-Selection framework with ten selector strategies, and the Distributional Alignment Score (DAS), a training-free, MMD-based quality estimator. All tracks are tested under shared domains, base models, training protocols, and downstream benchmarks, so methods are compared by their actual downstream impact.

🏗️ Data Construction

Data Construction

The Data Construction module md_to_qa converts Markdown books and long-form documents into structured supervision datasets for LLM fine-tuning. It targets book-to-SFT workflows where full content coverage, resumable execution, and quality control are required.

Dataset Outputs

The pipeline compiles source knowledge into three complementary supervision forms:

• Concept QA — Atomic, reusable knowledge such as definitions, categories, rules, mechanisms, purposes, and constraints.
• Process QA — Concise, grounded reasoning patterns including condition checking, rule application, causal explanation, comparison, exception handling, and step ordering.
• Case Application — Knowledge transfer into realistic, source-grounded scenarios where the model must analyze a situation and apply domain knowledge.

Pipeline Overview

1. Corpus Preparation — Build a manifest from a directory of Markdown files and split long documents into overlapping chunks.
2. Knowledge Cleaning — Clean and normalize chunks to remove boilerplate and improve semantic coherence.
3. QA Generation — Generate the three supervision forms above from each chunk via LLM-based operators.
4. Scoring & Filtering — Score generated QA pairs and filter out low-quality items.
5. Validation & Coverage Audit — Ensure every chunk reaches a final state (kept or skipped) and report coverage statistics.

The pipeline is resumable and tracks progress via chunk_status.jsonl, making it suitable for long-running batch jobs.

Implementation Layout

Path	Description
md_to_qa/DataFlow/	Core pipeline implementations, including chunking, cleaning, generation, scoring, and filtering operators.
md_to_qa/LLM/	Batch and domain processing scripts for large-scale data generation.
md_to_qa/SKILL/	Skill definitions, reference materials, and helper scripts for Markdown-to-QA conversion.
md_to_qa/Agent/	Agent prompts and task specifications for automated dataset construction.

Data Construction Skill

The underlying data-construction skill is also published as a standalone, reusable skill:

• Skill: data_construction_skill

You can reference or import this skill directly in compatible agent frameworks.

Training

Training is conducted using LlamaFactory. Base models include:

• Qwen2.5-7B
• Llama-3.1-8B

For each constructed dataset, we will use it to train both base models followed by a Dolly-15k training.

Please refer to Experiment.md for detailed configures we employed in our experiments.

Evaluation

The evaluation codes are in Data-Agent-Evaluation. You can use the script to run evaluation for the models trained in the last step. Please refer to README.md for instruction to use the script and Experiment.md for detailed configurations for evaluation.

🎯 Data Selection

The Data Selection framework (data-selection) selects high-utility subsets from large candidate SFT pools. It unifies random baselines, length and perplexity filters, embedding similarity, quality scoring, diversity algorithms, and LLM-as-a-judge methods behind a single Selector protocol, making it easy to compare selection strategies under identical training and evaluation conditions.

Average downstream score of each selection method across all domains and selection budgets (k). Higher is better.

Selector Overview

Selector	Strategy
RandomSelector	Random baseline
SourceBalancedRandomSelector	Source-balanced random sampling
LengthBasedSelector	Shortest / longest samples
PerplexityBasedSelector	Perplexity-based selection (low / high / mid)
EmbeddingSimilaritySelector	NEAR-style similarity to a target set
DeitaQualitySelector	Quality × complexity scoring
QualityScorerSelector	FineWeb-Edu / PairQual quality scoring
DiversityKCenterSelector	TSDS diversity K-Center
LLMAsSelector	LLM multi-dimensional scoring
CompositeSelector	Chain multiple selectors

Installation

cd data-selection
uv sync

Quick Start

from data_selection import RandomSelector
from data_selection.runner import run_selection

selector = RandomSelector(k=100, seed=42)
run_selection(
    selector,
    input_path="data/input.jsonl",
    output_path="data/output_random.jsonl",
)

run_selection also supports multiple selection budgets in one call. For score-based selectors, scoring runs once at max(k) and results are truncated for smaller k values:

run_selection(
    selector,
    input_path="data/input.jsonl",
    output_path="data/output_k{k}.jsonl",
    k=[1000, 10000, 100000],
)

Please refer to data-selection/README.md for a complete guide to all selectors, input formats, score caching, and development commands.

Evaluation

After selecting a subset, train a model on it with LlamaFactory and evaluate with the Data-Agent-Evaluation harness. Please refer to Experiment.md for selection results across finance, law, medicine, math, general, and science domains.

📊 Data Quality

A Python package for computing distributional distances (e.g., MMD) between datasets, designed for evaluating data preparation quality in LLM training pipelines.

Installation

This project uses uv for dependency management. To get started:

git clone https://github.com/haolpku/Data-Preparation-Bench.git
cd Data-Preparation-Bench
uv sync

To set up the development environment:

uv sync --extra dev
uv run pre-commit install

Before committing, format and lint the code:

uv run pre-commit run --all-files

Computing MMD Distance

The example script compute_mmd.py demonstrates how to compute MMD distance between two datasets using the vLLM OpenAI-compatible embedding API.

1. Start a vLLM embedding server (e.g., serving Qwen/Qwen3-Embedding-8B):

   vllm serve Qwen/Qwen3-Embedding-8B --runner pooling --trust-remote-code --max-model-len 40960 --served-model-name Qwen/Qwen3-Embedding-8B --dtype bfloat16 --seed 42

2. Configure the datasets in compute_mmd.py:

   from distflow.data.dataset import DistflowDataset
   from distflow.data.data_formatter import AlpacaFormatter, ShareGptFormatter
   
   dataset_1 = DistflowDataset(
       dataset_name="oda-math",
       data_path="OpenDataArena/ODA-Math-460k",
       load_type="datasets",
       formatter=AlpacaFormatter(
           user_key="question",
           assistant_key="response",
       ),
       data_size=5000,
       split="train",
       shuffle_seed=42,
   )
   
   dataset_2 = DistflowDataset(
       dataset_name="infinity-instruct",
       data_path="BAAI/Infinity-Instruct",
       load_type="datasets",
       formatter=ShareGptFormatter(
           conversations_key="conversations",
       ),
       data_size=5000,
       split="train",
       shuffle_seed=42,
   )

   Typically, you only need to update data_path with your HuggingFace dataset identifier and define a formatter that converts raw items to the required chat format. Two built-in formatters are available:

   • AlpacaFormatter: For datasets with separate user_key / assistant_key fields.
   • ShareGptFormatter: For datasets with a conversations field containing multi-turn messages.

3. Run the computation:

   uv run examples/compute_mmd.py

   To save results to a JSON file:

   uv run examples/compute_mmd.py --output results/

Please refer to Experiment.md for detailed configurations we employed in our experiments.

Training

Training is conducted using LlamaFactory. Base models include:

• Qwen2.5-7B
• Llama-3.1-8B
• Mistral-7B-v0.3

Please refer to Experiment.md for detailed configures we employed in our experiments.

Evaluation

The evaluation codes are in Data-Agent-Evaluation. You can use the script to run evaluation for the models trained in the last step. Please refer to README.md for instruction to use the script and Experiment.md for detailed configurations for evaluation.

Running Quality Benchmark

The example script run_benchmark.py shows how to evaluate a custom data-quality metric by measuring its correlation with downstream task accuracy.

1. Prepare datasets (same as above):

   Define one or more DistflowDataset objects with the appropriate formatter.

2. Provide accuracy values manually:

   The accuracy/ directory is not part of the repository, so you must supply your own accuracy mapping. The keys must exactly match the dataset_name field of each dataset:

   accuracies = {
       "dataflow": 0.25,
       "infinity-instruct": 0.30,
       "openr1": 0.45,
   }

3. Implement a metric class:

   Your metric class must implement score(dataset: DistflowDataset) -> list[MetricsResult]:

   class MyMetric:
       def score(self, dataset: DistflowDataset):
           # Compute metric for the dataset
           return [{"name": "my_metric", "value": 0.8, "meta": {}}]

4. Run the benchmark:

   uv run examples/run_benchmark.py

   The benchmark computes Pearson / Spearman correlation and a linear fit between your metric and the provided accuracies.

Please refer to Experiment.md for detailed accuracy results.

---

💬 Community