Update README: Refine experimental results, setup details, and L0 parser comparison

README.md

@@ -63,8 +63,8 @@ Experiments show that on the MiniCPM-1B architecture, ***UltraData-Math*** achie
 
 ***UltraData-Math*** has been applied to the mathematical pre-training of the [MiniCPM Series](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) models.
 
-- **[UltraData-Math-L1](https://huggingface.co/datasets/openbmb/UltraData-Math)**: Large-scale high-quality mathematical pre-training dataset, containing 159.4B tokens of web mathematical corpus. (**<-- you are here**)
-- **[UltraData-Math-L3](https://huggingface.co/datasets/openbmb/UltraData-Math-L3)**: High-quality synthetic mathematical dataset, containing 37.1B tokens of multi-format synthetic data (Q&A, multi-turn dialogues, knowledge textbooks, etc.).
+- **[UltraData-Math-L1](https://huggingface.co/datasets/openbmb/UltraData-Math)**: Large-scale high-quality mathematical pre-training dataset, containing 170.5B tokens of web mathematical corpus. (**<-- you are here**)
+- **[UltraData-Math-L3](https://huggingface.co/datasets/openbmb/UltraData-Math-L3)**: High-quality synthetic mathematical dataset, containing 88B tokens of multi-format synthetic data (Q&A, multi-turn dialogues, knowledge textbooks, etc.).
 
 ## 🏗️ Data Processing Pipeline
 
@@ -128,15 +128,29 @@ Natural web data is mostly declarative text. To enhance the model's instruction
 
 ## 📈 Experimental Results
 
-We used the **MiniCPM-1.2B** model architecture and **MiniCPM3-4B** tokenizer for experimental verification. Each experiment was conducted with a training volume of **100 billion Tokens**, allowing for comprehensive verification of data performance within a parameter range with controllable computational efficiency. We used the Lighteval library for model evaluation, and all evaluation metrics are based on **Zero-Shot** settings. Evaluation benchmarks include:
+We used the **MiniCPM-1.2B** model architecture and **MiniCPM3-4B** tokenizer for experimental verification. Each experiment was conducted with a training volume of **100 billion Tokens**, using the **Decay Verification** method (annealing from a 1.3T base model). We used the Lighteval library for model evaluation. Evaluation benchmarks include:
 
-- **Mathematical Reasoning:** GSM8K, MATH, R-Bench, Math-Bench
-- **Code Generation:** HumanEval, MBPP
-- **Comprehensive Knowledge:** MMLU, MMLU-STEM
+- **Mathematical Reasoning:** GSM8K (4-shot), MATH (4-shot), Math-Bench
+- **Code Generation:** HumanEval (0-shot), MBPP (3-shot)
+- **Comprehensive Knowledge:** MMLU (5-shot), MMLU-STEM (5-shot)
 
-### L0 Parser Comparison
+### 🔧 Experimental Setup
 
-Based on data from the same source, we re-parsed the raw HTML using different parsers and applied the same L1 cleaning operators to all parsed results. Each cleaned dataset was then used for independent training to directly compare the effects of parsing strategies:
+Following the **Decay Verification** method described in the paper, we evaluated the data quality by continuing the pre-training of a **MiniCPM-1.2B** base model (trained on 1.3T tokens).
+
+| Hyperparameter | Value | Description |
+| :--- | :--- | :--- |
+| **Base Model** | MiniCPM-1.2B | Pre-trained on 1.3T MiniCPM-4 corpus |
+| **Training Tokens** | ~100B | 20,000 steps |
+| **Data Mixture** | 30% Target / 70% Base | 30% UltraData-Math + 70% General Data |
+| **Sequence Length** | 4096 | |
+| **Global Batch Size** | 1280 | Micro batch size 10 |
+| **Learning Rate** | 7.5e-4 $\to$ 3.75e-5 | Exponential decay |
+| **Optimizer** | AdamW | with Maximal Update Parameterization (µP) |
+
+### Effectiveness of L0 Parsing Strategy
+
+To fairly compare different parsing strategies, we conducted experiments on a data subset sampled from the **2023-2024** distribution. We re-parsed the raw HTML from this source using different parsers and **applied the same L1 cleaning operators to all baselines**. This comparison demonstrates the **overall benefit of our L0 Parser + L1 Filtering pipeline** against other parsers under identical cleaning conditions.
 
 | Parser | Average | MMLU | GSM8K | HumanEval | math | mbpp_full | mmlu-stem |
 |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
@@ -147,6 +161,17 @@ Based on data from the same source, we re-parsed the raw HTML using different pa
 | magic-html + w3m | 41.29 | 51.23 | 51.63 | 26.83 | 26.58 | 45.02 | 46.45 |
 
 
+### Pipeline Effectiveness (L1 vs L2 vs L3)
+
+To validate the effectiveness of our L0-L3 hierarchical framework, we conducted ablation studies comparing models trained on different tiers of UltraData-Math. Unlike the L0 parser comparison above (which used a 2023-2024 subset), these results are based on the **full dataset**.
+
+| Dataset | Average | MMLU | ARC-E | ARC-C | BBH | CSQA | Hella. | OBQA | PIQA | SIQA | Wino. | Math | GSM8K | MBPP | HumanEval | CMMLU | C-Eval | Avg_ZH |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| **UltraData-Math-L1** | 48.39 | 51.41 | 54.50 | 37.29 | 37.75 | 60.44 | 58.02 | 41.60 | 74.21 | 41.71 | 57.14 | 27.78 | 54.66 | 44.71 | 29.88 | 51.28 | 51.89 | 51.59 |
+| **UltraData-Math-L2** | 48.59 | 50.93 | 55.20 | 36.95 | 39.27 | 60.20 | 57.52 | 39.80 | 74.48 | 44.73 | 57.77 | 29.20 | 52.92 | 44.50 | 32.32 | 51.13 | 50.55 | 50.84 |
+| **UltraData-Math-L3** | **52.23** | **51.67** | **59.79** | **38.98** | **43.62** | **61.18** | **58.27** | **57.00** | **74.76** | 43.35 | **59.04** | **37.02** | **61.79** | **49.27** | **32.93** | **52.87** | **54.08** | **53.48** |
+
+*Note: Results demonstrate that higher-tier data (L3) significantly boosts mathematical reasoning (MATH, GSM8K) and general capabilities.*
 
 ### Full Evaluation Results
 
@@ -166,6 +191,20 @@ We used a single dataset for independent training to directly compare the effect
 - **L3 Synthesis Layer**: [Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct), [Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B), [GLM-4.5](https://huggingface.co/zai-org/GLM-4.5)
 - **Seed Data**: [Nemotron-CC-Math](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1), [MegaMath](https://huggingface.co/datasets/LLM360/MegaMath), [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath)
 
+## 📖 Citation
+
+If you find **UltraData-Math** useful in your research, please consider citing:
+
+```bibtex
+@misc{ultradata-math,
+  title={UltraData-Math},
+  author={Chuyue Zhou and Hongya Lv and Xinle Lin and Yudong Wang and Jie Zhou and Hengyu Zhao and Junshao Guo and Xueren Zhang and Shuaikang Xue and Zhiyuan Liu},
+  year={2026},
+  url={https://huggingface.co/datasets/openbmb/UltraData-Math},
+  publisher={Hugging Face}
+}
+```
+
 ## 📜 License
 
 This project is licensed under the [Apache 2.0](./LICENSE) license.