ZhouChuYue commited on
Commit
3559ea7
·
1 Parent(s): bffeb3e

Update README: replace experiment tables with images, update benchmark list

Browse files
README.md CHANGED
@@ -162,32 +162,27 @@ ds = load_dataset("openbmb/UltraData-Math", "UltraData-Math-L3-Conversation-Synt
162
 
163
  We evaluated data quality using the **Decay Verification** method: continuing pre-training of a **MiniCPM-1.2B** base model (pre-trained on 1.3T tokens with **MiniCPM3-4B** tokenizer) with **~100B tokens** (30% target data + 70% general data). We used [OpenCompass](https://github.com/open-compass/opencompass) as our evaluation framework. Evaluation benchmarks include:
164
 
165
- - **Mathematical Reasoning:** GSM8K, MATH500, Math-Bench, R-Bench-Math
166
  - **Code Generation:** HumanEval, MBPP
167
- - **Comprehensive Knowledge:** MMLU, MMLU-STEM
 
168
 
169
  ### Effectiveness of L0 Parsing Strategy
170
 
171
  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. This comparison demonstrates the **effectiveness of our L0 Parser** against other parsers.
172
 
173
- | Parser | Average | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval |
174
- |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
175
- | **UltraData-Math-Parser (Ours)** | **43.44** | 51.41 | 46.76 | **28.72** | 54.97 | 47.10 | **31.71** |
176
- | trafilatura + w3m | 42.33 | 50.95 | 45.52 | 27.64 | 54.51 | **47.93** | 27.44 |
177
- | trafilatura | 42.44 | 51.42 | 46.62 | 28.08 | **56.03** | 45.64 | 26.83 |
178
- | Megamath | 42.32 | **51.46** | **46.81** | 26.04 | 54.06 | 45.64 | 29.88 |
179
- | magic-html + w3m | 41.29 | 51.23 | 46.45 | 26.58 | 51.63 | 45.02 | 26.83 |
180
 
181
 
182
  ### Pipeline Effectiveness (L1 vs L2 vs L3)
183
 
184
  To validate the effectiveness of our L0-L3 tiered 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**.
185
 
186
- | Dataset | Average | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval |
187
- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
188
- | **UltraData-Math-L1** | 42.31 | 51.41 | 45.44 | 27.78 | 54.66 | 44.71 | 29.88 |
189
- | **UltraData-Math-L2** | 42.57 | 50.93 | 45.52 | 29.20 | 52.92 | 44.50 | 32.32 |
190
- | **UltraData-Math-L3** | **46.44** | **51.67** | **45.93** | **37.02** | **61.79** | **49.27** | **32.93** |
191
 
192
  *Note: Results demonstrate that higher-tier data (L3) significantly boosts mathematical reasoning (MATH500, GSM8K) and general capabilities.*
193
 
@@ -195,13 +190,9 @@ To validate the effectiveness of our L0-L3 tiered framework, we conducted ablati
195
 
196
  To compare against existing public mathematical pre-training datasets, we trained models independently on each dataset using the same model architecture and training budget (~100B tokens). The baselines include [Nemotron-CC-Math](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1), [MegaMath-Web-Pro](https://huggingface.co/datasets/LLM360/MegaMath), and [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath). All models are evaluated under identical conditions for a fair comparison:
197
 
198
- | Model | Average | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval | R-Bench-Math | Math-Bench |
199
- |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
200
- | **UltraData-Math (Ours)** | **43.79** | 51.67 | 45.93 | **37.02** | **61.79** | **49.27** | 32.93 | 23.38 | **48.33** |
201
- | Nemotron-cc 4plus mind | 43.45 | 52.09 | 45.99 | 35.96 | 59.97 | 48.03 | 34.76 | **23.51** | 47.25 |
202
- | Nemotron-cc 4plus | 42.62 | 51.96 | 45.67 | 33.40 | 58.45 | 46.47 | **35.37** | 22.74 | 46.92 |
203
- | MegaMath-Web-Pro | 41.38 | **53.16** | **47.15** | 32.12 | 56.71 | 47.10 | 31.71 | 21.23 | 41.83 |
204
- | FineMath-4+ | 40.51 | 50.90 | 44.98 | 29.84 | 56.25 | 48.96 | 29.88 | 18.93 | 44.33 |
205
 
206
  ## ❤️ Acknowledgements
207
 
 
162
 
163
  We evaluated data quality using the **Decay Verification** method: continuing pre-training of a **MiniCPM-1.2B** base model (pre-trained on 1.3T tokens with **MiniCPM3-4B** tokenizer) with **~100B tokens** (30% target data + 70% general data). We used [OpenCompass](https://github.com/open-compass/opencompass) as our evaluation framework. Evaluation benchmarks include:
164
 
165
+ - **Mathematical Reasoning:** MATH500, GSM8K, Math-Bench, R-Bench-Math
166
  - **Code Generation:** HumanEval, MBPP
167
+ - **English:** MMLU, ARC-E, ARC-C, BBH, CSQA, HellaSwag, OBQA, PIQA, SIQA, WinoGrande
168
+ - **Chinese:** CMMLU, C-Eval
169
 
170
  ### Effectiveness of L0 Parsing Strategy
171
 
172
  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. This comparison demonstrates the **effectiveness of our L0 Parser** against other parsers.
173
 
174
+ <div align="center">
175
+ <img src="assets/ultradata-math-l0-parser-comparison.png" width="700"/>
176
+ </div>
 
 
 
 
177
 
178
 
179
  ### Pipeline Effectiveness (L1 vs L2 vs L3)
180
 
181
  To validate the effectiveness of our L0-L3 tiered 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**.
182
 
183
+ <div align="center">
184
+ <img src="assets/ultradata-math-l1l2l3-comparison.png" width="700"/>
185
+ </div>
 
 
186
 
187
  *Note: Results demonstrate that higher-tier data (L3) significantly boosts mathematical reasoning (MATH500, GSM8K) and general capabilities.*
188
 
 
190
 
191
  To compare against existing public mathematical pre-training datasets, we trained models independently on each dataset using the same model architecture and training budget (~100B tokens). The baselines include [Nemotron-CC-Math](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1), [MegaMath-Web-Pro](https://huggingface.co/datasets/LLM360/MegaMath), and [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath). All models are evaluated under identical conditions for a fair comparison:
192
 
193
+ <div align="center">
194
+ <img src="assets/ultradata-math-full-comparison.png" width="700"/>
195
+ </div>
 
 
 
 
196
 
197
  ## ❤️ Acknowledgements
198
 
README_ZH.md CHANGED
@@ -18,7 +18,7 @@
18
  - **数据质量层面**:现有数据集普遍缺乏系统的质量分级机制,高价值数学内容与低质噪声混杂。
19
  - **数据多样性层面**:主流数据集多源自教科书或竞赛题库,缺少真实网页中的数学讨论与应用场景;合成数据格式单一,难以覆盖多轮对话、多风格表达等多样化需求。
20
 
21
- 针对上述问题,我们提出 ***UltraData-Math***——一个面向数学推理任务的大规模高质量预训练数据集。本数据集基于 [UltraData](https://huggingface.co/collections/openbmb/ultradata) 的 L0-L4 分级数据管理框架开发,包含四个递进层级:
22
 
23
  - **L0 原始数据层**:基于 *magic-html* 开发数学解析器,结合 *w3m* 布局保持渲染与多级回退策略,将 MathML、KaTeX、AsciiMath 标准化为 LaTeX 格式。
24
  - **L1 过滤数据层**:通过启发式规则清洗噪声并进行文档级去重。
@@ -35,7 +35,7 @@
35
 
36
  ## 🏗️ 数据处理流水线
37
 
38
- 为突破现有数学数据集在质量与多样性上的局限,我们建立了一套以"数学内容完整性"和"信息密度"为核心的精细化分级标准。***UltraData-Math*** 采用了 [UltraData](https://huggingface.co/collections/openbmb/ultradata) 论文提出的 **L0-L4 分级数据管理框架**,通过标准化的层级定义,实现数学数据资产的有序管理与高效流转。每一级都代表了更高的数据纯度与数学价值,同时也对应着更精细的加工程度。
39
 
40
  <div align="center">
41
  <img src="assets/ultradata-math-pipeline.png" width="900"/>
@@ -123,32 +123,27 @@ ds = load_dataset("openbmb/UltraData-Math", "UltraData-Math-L3-Conversation-Synt
123
 
124
  我们使用 **衰减验证(Decay Verification)** 方法评估数据质量:在 **MiniCPM-1.2B** 基座模型(使用 **MiniCPM3-4B** 分词器,预训练 1.3T tokens)上继续训练 **~100B tokens**(30% 目标数据 + 70% 通用数据)。我们使用 [OpenCompass](https://github.com/open-compass/opencompass) 作为评估框架。评估基准包括:
125
 
126
- - **数学推理:** GSM8K、MATH500、Math-Bench、R-Bench-Math
127
  - **代码生成:** HumanEval、MBPP
128
- - **综合知识:** MMLU、MMLU-STEM
 
129
 
130
  ### L0 解析策略有效性
131
 
132
  为公平对比不同解析策略,我们在 **2023-2024** 年分布的数据子集上进行实验。我们使用不同解析器重新解析原始 HTML。该对比展示了我们 **L0 解析器的有效性**。
133
 
134
- | 解析器 | 平均分 | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval |
135
- |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
136
- | **UltraData-Math-Parser (Ours)** | **43.44** | 51.41 | 46.76 | **28.72** | 54.97 | 47.10 | **31.71** |
137
- | trafilatura + w3m | 42.33 | 50.95 | 45.52 | 27.64 | 54.51 | **47.93** | 27.44 |
138
- | trafilatura | 42.44 | 51.42 | 46.62 | 28.08 | **56.03** | 45.64 | 26.83 |
139
- | Megamath | 42.32 | **51.46** | **46.81** | 26.04 | 54.06 | 45.64 | 29.88 |
140
- | magic-html + w3m | 41.29 | 51.23 | 46.45 | 26.58 | 51.63 | 45.02 | 26.83 |
141
 
142
 
143
  ### 流水线有效性(L1 vs L2 vs L3)
144
 
145
  为验证 L0-L3 分级框架的有效性,我们对使用不同层级 UltraData-Math 训练的模型进行了消融实验。与上文 L0 解析器对比(使用 2023-2024 子集)不同,以下结果基于**全量数据集**。
146
 
147
- | 数据集 | 平均分 | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval |
148
- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
149
- | **UltraData-Math-L1** | 42.31 | 51.41 | 45.44 | 27.78 | 54.66 | 44.71 | 29.88 |
150
- | **UltraData-Math-L2** | 42.57 | 50.93 | 45.52 | 29.20 | 52.92 | 44.50 | 32.32 |
151
- | **UltraData-Math-L3** | **46.44** | **51.67** | **45.93** | **37.02** | **61.79** | **49.27** | **32.93** |
152
 
153
  *注:结果表明,更高层级的数据(L3)显著提升了数学推理能力(MATH500、GSM8K)及通用能力。*
154
 
@@ -156,13 +151,9 @@ ds = load_dataset("openbmb/UltraData-Math", "UltraData-Math-L3-Conversation-Synt
156
 
157
  为与现有公开数学预训练数据集进行对比,我们使用相同的模型架构和训练预算(~100B tokens)在每个数据集上独立训练模型。基线包括 [Nemotron-CC-Math](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1)、[MegaMath-Web-Pro](https://huggingface.co/datasets/LLM360/MegaMath) 和 [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath)。所有模型在相同条件下评估以确保公平对比:
158
 
159
- | 模型 | 平均分 | MMLU | MMLU-STEM | MATH500 | GSM8K | MBPP | HumanEval | R-Bench-Math | Math-Bench |
160
- |:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
161
- | **UltraData-Math (Ours)** | **43.79** | 51.67 | 45.93 | **37.02** | **61.79** | **49.27** | 32.93 | 23.38 | **48.33** |
162
- | Nemotron-cc 4plus mind | 43.45 | 52.09 | 45.99 | 35.96 | 59.97 | 48.03 | 34.76 | **23.51** | 47.25 |
163
- | Nemotron-cc 4plus | 42.62 | 51.96 | 45.67 | 33.40 | 58.45 | 46.47 | **35.37** | 22.74 | 46.92 |
164
- | MegaMath-Web-Pro | 41.38 | **53.16** | **47.15** | 32.12 | 56.71 | 47.10 | 31.71 | 21.23 | 41.83 |
165
- | FineMath-4+ | 40.51 | 50.90 | 44.98 | 29.84 | 56.25 | 48.96 | 29.88 | 18.93 | 44.33 |
166
 
167
  ## ❤️ 致谢
168
 
 
18
  - **数据质量层面**:现有数据集普遍缺乏系统的质量分级机制,高价值数学内容与低质噪声混杂。
19
  - **数据多样性层面**:主流数据集多源自教科书或竞赛题库,缺少真实网页中的数学讨论与应用场景;合成数据格式单一,难以覆盖多轮对话、多风格表达等多样化需求。
20
 
21
+ 针对上述问题,我们提出 ***UltraData-Math***——一个面向数学推理任务的大规模高质量预训练数据集。本数据集基于 [UltraData](https://ultradata.openbmb.cn/blog/position-paper) 的 L0-L4 分级数据管理框架开发,包含四个递进层级:
22
 
23
  - **L0 原始数据层**:基于 *magic-html* 开发数学解析器,结合 *w3m* 布局保持渲染与多级回退策略,将 MathML、KaTeX、AsciiMath 标准化为 LaTeX 格式。
24
  - **L1 过滤数据层**:通过启发式规则清洗噪声并进行文档级去重。
 
35
 
36
  ## 🏗️ 数据处理流水线
37
 
38
+ 为突破现有数学数据集在质量与多样性上的局限,我们建立了一套以"数学内容完整性"和"信息密度"为核心的精细化分级标准。***UltraData-Math*** 采用了 [UltraData](https://ultradata.openbmb.cn/blog/position-paper) 论文提出的 **L0-L4 分级数据管理框架**,通过标准化的层级定义,实现数学数据资产的有序管理与高效流转。每一级都代表了更高的数据纯度与数学价值,同时也对应着更精细的加工程度。
39
 
40
  <div align="center">
41
  <img src="assets/ultradata-math-pipeline.png" width="900"/>
 
123
 
124
  我们使用 **衰减验证(Decay Verification)** 方法评估数据质量:在 **MiniCPM-1.2B** 基座模型(使用 **MiniCPM3-4B** 分词器,预训练 1.3T tokens)上继续训练 **~100B tokens**(30% 目标数据 + 70% 通用数据)。我们使用 [OpenCompass](https://github.com/open-compass/opencompass) 作为评估框架。评估基准包括:
125
 
126
+ - **数学推理:** MATH500、GSM8K、Math-Bench、R-Bench-Math
127
  - **代码生成:** HumanEval、MBPP
128
+ - **英文:** MMLU、ARC-E、ARC-C、BBH、CSQA、HellaSwag、OBQA、PIQA、SIQA、WinoGrande
129
+ - **中文:** CMMLU、C-Eval
130
 
131
  ### L0 解析策略有效性
132
 
133
  为公平对比不同解析策略,我们在 **2023-2024** 年分布的数据子集上进行实验。我们使用不同解析器重新解析原始 HTML。该对比展示了我们 **L0 解析器的有效性**。
134
 
135
+ <div align="center">
136
+ <img src="assets/ultradata-math-l0-parser-comparison.png" width="700"/>
137
+ </div>
 
 
 
 
138
 
139
 
140
  ### 流水线有效性(L1 vs L2 vs L3)
141
 
142
  为验证 L0-L3 分级框架的有效性,我们对使用不同层级 UltraData-Math 训练的模型进行了消融实验。与上文 L0 解析器对比(使用 2023-2024 子集)不同,以下结果基于**全量数据集**。
143
 
144
+ <div align="center">
145
+ <img src="assets/ultradata-math-l1l2l3-comparison.png" width="700"/>
146
+ </div>
 
 
147
 
148
  *注:结果表明,更高层级的数据(L3)显著提升了数学推理能力(MATH500、GSM8K)及通用能力。*
149
 
 
151
 
152
  为与现有公开数学预训练数据集进行对比,我们使用相同的模型架构和训练预算(~100B tokens)在每个数据集上独立训练模型。基线包括 [Nemotron-CC-Math](https://huggingface.co/datasets/nvidia/Nemotron-CC-Math-v1)、[MegaMath-Web-Pro](https://huggingface.co/datasets/LLM360/MegaMath) 和 [FineMath](https://huggingface.co/datasets/HuggingFaceTB/finemath)。所有模型在相同条件下评估以确保公平对比:
153
 
154
+ <div align="center">
155
+ <img src="assets/ultradata-math-full-comparison.png" width="700"/>
156
+ </div>
 
 
 
 
157
 
158
  ## ❤️ 致谢
159
 
assets/ultradata-math-full-comparison.png ADDED

Git LFS Details

  • SHA256: a6bb63525dd382940346b1c8cf2a4a463d89ce9b196f440d2c20bd047d5585dd
  • Pointer size: 130 Bytes
  • Size of remote file: 41.8 kB
assets/ultradata-math-l0-parser-comparison.png ADDED

Git LFS Details

  • SHA256: e9ff6e5d1bbcc67686d665bf398adabf6cb203b56329fb267bdff4fa18ffac28
  • Pointer size: 130 Bytes
  • Size of remote file: 34.2 kB
assets/ultradata-math-l1l2l3-comparison.png ADDED

Git LFS Details

  • SHA256: 346ccd7b1f875fccc20f2bf901c4005a29ae0e7823d8a56ba8117f053ffe8f72
  • Pointer size: 131 Bytes
  • Size of remote file: 174 kB