DataMan: Data Manager for Pre-training Large Language Models

This is the official repository for our ICLR'25 paper DataMan: Data Manager for Pre-training Large Language Models. It provides code to:

1. Collect fine-tuning data—document quality ratings and domain types—via a SuperLLM;
2. Train and evaluate DataMan models;
3. Annotate text quality and domain by DataMan;
4. Select and mix data by DataMan Annotation;
5. Train language models;
6. Evaluate language models;
7. Reproduce the paper’s analyses.

Note: This repository is solely developed and maintained by me. If you find it helpful, feel free to follow and give a ⭐ to support my hard work. Due to limited bandwidth, I may not be able to respond to every issue quickly.

Pipeline of the Sample-with-DataMan framework.

Datasets

The fine-tuning data are subject to license and confidentiality agreements; releasing them would vithereforeolate company policy. Consequently, we share only the preprocessing/analysis code and the public pre-training dataset SlimPajama. The paper’s DataPajama set contains 447B tokens and is too large to release for now, users can reconstruct it with the supplied code and DataMan models.

Models

DataMan Models

🔥 Four variants (Dense/MoE × EN/ZH) are now on the HuggingFace Hub. 🔥

• Dense: fine-tuned from Qwen2-1.5B
• MoE: fine-tuned from Qwen1.5-MoE-A2.7B
• English/Chinese: The two versions differ only in the language of the fine-tuning corpus.

Variant	Link
English Dense	RuPeng/DataMan-1.5B-EN
Chinese Dense	RuPeng/DataMan-1.5B-ZH
English MoE	RuPeng/DataMan-MoE-A2.7B-EN
Chinese MoE	RuPeng/DataMan-MoE-A2.7B-ZH

Language Models

Using different data selection methods, we select a 30B token subset from either 447B-token DataPajama or 296B-token DataChineseWebText and train a randomly initialized either Sheared-Llama-1.3B or Qwen2.5-1.5B language model for one epoch in a randomly shuffled order.

Most baseline models referenced in the paper can be downloaded from princeton-nlp/QuRating. Our trained models have not yet passed the company’s open-source disclosure process (in-progress).

Experiments (English task shown; the process is identical for Chinese)

1. Installing the Repo and Environment

Clone this repo and create a new environment based on python 3.11.7. Install the requirements in the following order:

pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2
pip install -r requirements.txt

My Conda environment configuration is available in config.yml for reference.

2. Collecting fine-tuning Data

The Python file sft/prepare_data.py, together with the shell scripts listed below, produces quality ratings and domain types for each fine-tuning document using GPT-4. All scripts operate on JSON-formatted documents, and the prompt templates provides in sft/prepare_data.py are identical to those used in the paper.

(1) Build requests: Generate input files required by GPT.

bash scripts/sft/en/prepare_data/prepare_gpt_request_4k.sh
bash scripts/sft/en/prepare_data/prepare_gpt_request_4k_suppl.sh

(2) Obtain responses: Query the OpenAI API with the generated requests (the calling script is withheld pending compliance review).

(3) Post-process responses: Parse GPT outputs and normalize the format.

bash scripts/sft/en/prepare_data/process_gpt_response_4k.sh
bash scripts/sft/en/prepare_data/process_gpt_response_4k_suppl.sh

(4) Create DataMan fine-tuning datasets:

All-rating: 14 quality ratings + a domain type

bash scripts/sft/en/prepare_data/all_rating/prepare_qwen_request_suppl_4k_bal_q.sh

Score-Only: 14 quality ratings

bash scripts/sft/en/prepare_data/score_only/prepare_qwen_request_score_4k_bal_q.sh

Domain-Only: a domain type

bash scripts/sft/en/prepare_data/domain_only/prepare_qwen_request_detText_domain_4k_bal_q.sh

3. Training & evaluating DataMan models

The main python file is in sft/finetune_qwen2.py and sft/modeling_qwen2.py. The shell scripts below launch training and evaluation for various configurations; hyper-parameters can be edited directly in each script.

(1) All-rating (Dense) variant:

bash scripts/sft/en/all_rating/suppl_models/finetune_qwen2_1_5B_4k_suppl_bal_q4.sh
bash scripts/sft/en/all_rating/suppl_models/evaluate_qwen2_1_5B_4k_suppl_bal_q4.sh

(2) Score-Only (Dense) variant:

bash scripts/sft/en/score_only/modelscale_models/finetune_qwen2_score_1_5B_4k_bal_q4.sh
bash scripts/sft/en/score_only/modelscale_models/evaluate_qwen2_score_1_5B_4k_bal_q4.sh

(3) Domain-Only (Dense) variant:

bash scripts/sft/en/domain_only/modelscale_models/finetune_qwen2_detText_domain_1_5B_4k_bal_q4.sh
bash scripts/sft/en/domain_only/modelscale_models/evaluate_qwen2_detText_domain_1_5B_4k_bal_q4.sh

(4) All-rating (MoE) variant:

bash scripts/sft/en/all_rating/moe_models/finetune_qwen1_5_moe_2_7B_4k_bal_q4_multinodes.sh
bash scripts/sft/en/all_rating/moe_models/evaluate_qwen1_5_moe_2_7B_4k_bal_q4.sh

4. Annotating quality & domain with DataMan

Taking SlimPajama as an example, we demonstrate the full workflow of annotating document quality and domain with DataMan. If your data are already pre-processed, skip directly to Step (3).

(1) Acquire the Slimpajama dataset

First, we can download the dataset using one of the three methods in utils/hf_download.py. Then, we decompress the .zst archives with scripts/data_tools/en/get_slimpajama_org/unzip_slimpajama_zst.sh. Next, we merge chunk files for each split via scripts/data_tools/en/get_slimpajama_org/cat_{train,validation,test}_chunk.sh. Finally, we werify file integrity by scripts/data_tools/en/get_slimpajama_org/check_chunk.sh.

(2) Tokenize the SlimPajama dataset

a. data_tools/tokenize_dataset_concatenation.py: tokenize + truncate long sentences + drop short ones

bash scripts/data_tools/en/tokenize_dataset/tokenize_dataset_dsw.sh

b. data_tools/tokenize_dataset.py: tokenize + truncate long sentences + concatenate short ones

bash scripts/data_tools/en/tokenize_dataset/tokenize_dataset_concat.sh

To further split the tokenized data into more shards, you can use data_tools/split_hf_shard.py:

bash scripts/data_tools/en/tokenize_dataset/split_hf_shard_dsw.sh

(3) Annotating the SlimPajama dataset with DataMan (⭐ Core Usage ⭐)

a. For JSON files: data_tools/dataman_annotate.py

bash scripts/data_tools/en/dataman_annotate/Slimpajama-627B/dense/dataman_annotate_all_rating_dsw.sh

b. For HuggingFace files: data_tools/dataman_annotate_hf.py

bash scripts/data_tools/en/dataman_annotate/Slimpajama-627B/dense/dataman_annotate_hf_all_rating_dsw.sh

5. Selecting & mixing data via annotations (⭐ Core Usage ⭐)

Continuing with SlimPajama, we demonstrate how to employ the core python file /DataMan/data_tools/select_subset.py to select and mixup the dataset based on the annotated quality and domain labels. Ready-made sampling recipes are provided for various scenarios:

a. Individual quality metrics (14 in total): scripts/data_tools/en/select_subset/dense/individual_quality_criteria

b. Composite metrics (Overall_score and quality_criteria_mix): scripts/data_tools/en/select_subset/dense/overall_score, scripts/data_tools/en/select_subset/dense/quality_criteria_mix

c. Uniform sampling: scripts/data_tools/en/select_subset/dense/uniform

d. Domain-specific sampling: scripts/data_tools/en/select_subset/dense/specific_domain

e. 60B-token subset: scripts/data_tools/en/select_subset/dense/60B

6. Training language models

All pre-training python files are located in DataMan/pretrain and all shell scripts are organised as follows:

a. base model: Sheared-LLaMA-1.3B (English), 30 B tokens: scripts/pretrain/en/dense/30B

b. base model: Sheared-LLaMA-1.3B (English), 60 B tokens: scripts/pretrain/en/dense/60B

c. base model: Sheared-LLaMA-1.3B (English), domain-specific pre-training (finance, law, medicine): scripts/pretrain/en/dense/cpt

d. base model: Qwen 2.5-1.5B (Chinese), 30 B tokens: scripts/pretrain/zh/moe/30B

7. Evaluating language models

We evaluate the trained models on three aspects: in-context learning (ICL), perplexity (PPL), and instruction-following win rate (requires an OpenAI API key). Source code for these evaluations resides in DataMan/eval, ICL: eval/lm-evaluation-harness, PPL: eval/loglikelihood.py, instruction following: eval/instruction_tuning.

Evaluation shell scripts: i)-English tasks: DataMan/scripts/eval/en/dense/30B (single-GPU and multi-GPU versions available); ii)-Chinese tasks: DataMan/scripts/eval/zh/moe/30B.

8. Analyses

(1) Data inspection: examine examples of original documents from each source under Dataman’s quality criteria and ratings. Shell Script: analysis/en/data_inspection.py, Python Code: scripts/analysis/en/data_inspection.sh

(2) DataMan pointwise rating vs. Qurating pairwise rating. Shell Script: analysis/sorted_field.py. Annotation Output: analysis/DataMan_Pointwise_vs_Qurating_Pairwise_Ranking.

(3) Complementarity nature of quality criteria: compute Pearson correlation heatmap between 14 quality criteria in the fine-tuning dataset. Python Code: analysis/overallscore_vs_other_criteria.py

(4) PPL anomalies: locate and explain documents with perplexity outliers. Python Code: analysis/ppl_anomalies.py

(5) Distribution of quality ratings: plot the distribution of quality ratings across different sources in DataPajama. Python Code: analysis/quality_rating_dist_dense.py

(6) Dataset statistics for EN / ZH pre-training, fine-tuning dataset: the number and proportion of documents categorized by domain, Overall Score, and source • EN pre-train: analysis/statistic_sft_data.py • ZH pre-train: analysis/statistic_chinesewebtext2_data.py • Fine-tune: analysis/statistic_sft_data.py

Citation

If you use the code in your research, please cite:

@article{peng2025dataman,
  title={Dataman: Data manager for pre-training large language models},
  author={Peng, Ru and Yang, Kexin and Zeng, Yawen and Lin, Junyang and Liu, Dayiheng and Zhao, Junbo},
  journal={arXiv preprint arXiv:2502.19363},
  year={2025}
}