rntc commited on
Commit
b16bc40
·
verified ·
1 Parent(s): 5e4cf3f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +206 -0
README.md CHANGED
@@ -47,4 +47,210 @@ configs:
47
  path: data/commercial-*
48
  - split: noncommercial
49
  path: data/noncommercial-*
 
50
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47
  path: data/commercial-*
48
  - split: noncommercial
49
  path: data/noncommercial-*
50
+ pretty_name: Biomed-Enriched
51
  ---
52
+ ![Biomed-Enriched](logo.png)
53
+
54
+ # Biomed-Enriched: A Biomedical Dataset Enriched with LLMs for Pretraining and Extracting Rare and Hidden Content
55
+
56
+ ### Dataset Authors
57
+
58
+ **Rian Touchent, Nathan Godey & Eric de la Clergerie**
59
+ *Sorbonne Université, INRIA Paris*
60
+
61
+ ### Overview
62
+
63
+ Biomed-Enriched is a PubMed-derived dataset created using a two-stage annotation process. Initially, [Llama 3.1 70B Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) annotated 400K paragraphs for document type, domain, and educational quality. These annotations were then used to fine-tune a smaller model, which propagated the labels across the entire PubMed Central Open Access corpus. This process yielded 2M clinical case paragraphs, with over 450K high-quality paragraphs licensed for commercial use. This dataset provides a large-scale, openly available alternative to private clinical text. In continual pre-training experiments with OLMo2, curated subsets showed targeted improvements: clinical upsampling boosted MMLU ProfMed scores by ~5%, and educational quality filtering improved MedQA and MedMCQA by ~1%. Combining these methods achieved similar performance compared to standard continual-pretraining with just one-third of the training tokens, highlighting the potential for more efficient biomedical pretraining.
64
+
65
+ The dataset is structured into two primary splits:
66
+
67
+ * **Commercial**
68
+ * **Non-Commercial**
69
+
70
+ ## Dataset Structure
71
+
72
+ ### Commercial Split
73
+
74
+ * **text**: Textual content of the paragraphs.
75
+ * **path**: Precise XML path referencing original paragraph locations.
76
+ * **license\_url**: URL linking to the license.
77
+ * **authors**: Comprehensive list of authors per paragraph for proper attribution compliance.
78
+
79
+ ### Non-Commercial Split
80
+
81
+ * **path**: Precise XML path referencing original paragraph locations.
82
+ * **license\_url**: URL linking to the license.
83
+ * **authors**: Comprehensive list of authors per paragraph for proper attribution compliance.
84
+
85
+ > **Note:** The non-commercial split does not contain text data due to licensing restrictions. However, we provide scripts to populate the `text` field from a local PMC Open Access XML dump. See below for installation and usage instructions.
86
+
87
+ ```bash
88
+ pip install biomed-enriched
89
+ ```
90
+
91
+ ### With Python
92
+ ```python
93
+ from biomed_enriched import populate
94
+
95
+ DATASET_DIR = "/path/to/biomed-enriched" # input dataset
96
+ PMC_XML_ROOT = "/path/to/pmc/non-comm/xml" # PMC XML dump
97
+ OUTPUT_DIR = "/path/to/populated-biomed-enriched" # drop arg to overwrite in-place
98
+
99
+ populate(DATASET_DIR, PMC_XML_ROOT, output_path=OUTPUT_DIR, splits="noncommercial", num_proc=1)
100
+ ```
101
+ The call overwrites the dataset in-place, adding a new `text` column as the third column (after `article_id`, `path`).
102
+
103
+ ### With CLI
104
+ ```bash
105
+ biomed-enriched \
106
+ --input /path/to/biomed-enriched \
107
+ --xml-root /path/to/pmc/non-comm/xml \
108
+ --num-proc 8
109
+ ```
110
+ Add `--output DIR` if you prefer writing to a new directory instead of overwriting.
111
+
112
+ ## Annotation Process
113
+
114
+ The dataset was created using a two-stage annotation framework:
115
+
116
+ 1. **Initial Annotation by Large Language Model**:
117
+
118
+ * Annotated a subset of paragraphs for the following categories:
119
+
120
+ - **Document Type**: Categorizes the structure and purpose of the content.
121
+ - **Clinical Case**: Detailed report of symptoms, diagnosis, treatment, and follow-up of individual patients.
122
+ - **Study**: Research paragraph with methods, results, and discussion of experiments or observations.
123
+ - **Review**: Summary or synthesis of current knowledge on a specific topic.
124
+ - **Other**: Content not fitting above categories (editorials, commentaries, policy paragraphs).
125
+ - **Domain**: Identifies the subject area focus.
126
+ - **Clinical**: Content relating to patient care, clinical trials, case reports, or practice guidelines.
127
+ - **Biomedical**: Scientific aspects of medicine and biology.
128
+ - **Other**: Content mentioning biomedical topics but focusing on administrative, policy, or general communications.
129
+ - **Educational Quality**: Assesses pedagogical value for college-level biomedical learning on a scale from 1 (minimal value) to 5 (exceptional value) inspired by FineWeb-edu.
130
+ - **Score 1**: Basic information relevant to biomedical topics, may contain irrelevant content.
131
+ - **Score 2**: Addresses biomedical education elements but with limitations in coherence or depth.
132
+ - **Score 3**: Appropriate for college-level curricula, introduces key concepts with reasonable coherence.
133
+ - **Score 4**: Highly relevant educational content with clear writing style, minimal irrelevant information.
134
+ - **Score 5**: Outstanding educational value, detailed reasoning with profound insights for college-level learning.
135
+
136
+ 2. **Annotation Scaling via Model Distillation**:
137
+
138
+ * These annotations were distilled into a XLM-RoBERTa-base model, enabling scalability to the entire PMC dataset.
139
+
140
+ ## Annotation Statistics
141
+
142
+ Here is the distribution of educational scores per domain:
143
+
144
+ | Educational Score | Biomedical (n=116 221 134) | Clinical (n=2 182 784) | Other (n=15 213 051) |
145
+ | :---------------: | :------------------------: | :--------------------: | :------------------: |
146
+ | 1 | 1.8 % | 6.0 % | 60.1 % |
147
+ | 2 | 9.4 % | 23.4 % | 29.4 % |
148
+ | 3 | 10.9 % | 26.6 % | 8.3 % |
149
+ | 4 | 75.3 % | 44.0 % | 2.1 % |
150
+ | 5 | 2.6 % | – | – |
151
+
152
+ Here is the distribution of educational scores per document type:
153
+
154
+ | Educational Score | Study (n=100 387 809) | Review (n=6 811 226) | Clinical case (n=2 122 403) | Other (n=24 295 531) |
155
+ |:-----------------:|:---------------------:|:--------------------:|:---------------------------:|:--------------------:|
156
+ | 1 | 0.7 % | 0.3 % | 4.0 % | 43.4 % |
157
+ | 2 | 7.9 % | 1.6 % | 14.4 % | 30.9 % |
158
+ | 3 | 10.0 % | 5.0 % | 24.6 % | 14.6 % |
159
+ | 4 | 78.7 % | 86.9 % | 57.0 % | 11.1 % |
160
+ | 5 | 2.6 % | 6.1 % | 0.0 % | 0.0 % |
161
+
162
+ ### Language Distribution
163
+
164
+
165
+ | Language | Articles | Paragraphs | Clinical Case Paragraphs | % Clinical Cases |
166
+ |----------|----------|------------|-------------------------|-----------------|
167
+ | en | 4,113,275| 131,579,445| 2,113,185 | 1.61 |
168
+ | es | 4,339 | 181,779 | 1,235 | 0.68 |
169
+ | zh-cn | 3,649 | 59,719 | 0 | 0.00 |
170
+ | fr | 3,410 | 173,325 | 2,586 | 1.49 |
171
+ | de | 2,976 | 248,608 | 51 | 0.02 |
172
+ | it | 2,708 | 274,819 | 521 | 0.19 |
173
+ | pt | 934 | 85,242 | 4,540 | 5.33 |
174
+ | ko | 636 | 25,535 | 0 | 0.00 |
175
+ | ru | 222 | 10,553 | 0 | 0.00 |
176
+ | id | 189 | 91,865 | 15 | 0.02 |
177
+
178
+
179
+ ## Key Applications
180
+
181
+ * Improve efficiency in biomedical pretraining by focusing on high-quality, specific content.
182
+ * Create new biomedical subsets tailored to specific research needs based on document type and domain.
183
+
184
+ ## Evaluation
185
+
186
+ Our evaluation focuses on isolating the effects of data curation rather than pursuing state-of-the-art scores on benchmarks. A more powerful foundation model would likely yield higher absolute scores but would obscure the precise impact of our dataset. We therefore selected `OLMo2-7B-stage1` as our foundation model, as this intermediate checkpoint provides strong baseline capabilities while allowing for a clear attribution of performance gains to our enrichment strategies. This model has already developed strong language modeling capabilities but precedes the knowledge-intensive tuning of stage 2, providing an ideal balance without the risk of catastrophic forgetting of instruction-following abilities during domain adaptation. Notably, the data mix used in phase 1 includes DCLM, a dataset filtered from web data using a classifier trained on instruction data, which gives OLMo2-7B relatively strong question-answering capabilities even after stage 1.
187
+
188
+ Each Biomed-Enriched variant was trained for exactly 33.6 billion tokens using identical hyperparameters. We follow the annealing strategy of OLMo2 used in the mid-training phase. By maintaining strict parameter parity across experiments, we created a controlled environment focused solely on measuring the effectiveness of different data curation strategies.
189
+
190
+ These experiments are designed to illustrate how our granular annotations enable targeted improvements in model capabilities. For instance, by specifically upsampling clinical content (`BE-Clinical` and `BE-ClinicalCase` variants), we expect to see a notable increase in performance on the MMLU Professional Medicine benchmark, underscoring the dataset's potential for developing specialized models.
191
+
192
+ The following variants were created for this evaluation:
193
+ * **BE-Base:** The complete unmodified PMC Open Access Subset serving as baseline.
194
+ * **BE-Educational:** Preserves all articles but removes paragraphs with educational quality scores below 3.
195
+ * **BE-Clinical:** Replicates articles with predominantly clinical domain content 10�� in the training mix.
196
+ * **BE-ClinicalCase:** Replicates articles containing at least one clinical case paragraph 10× to increase exposure to clinical narratives.
197
+ * **BE-Prefix:** Prefixes each paragraph with its predicted annotations to allow modeling of metadata-content relationships.
198
+ * **BE-French:** Upsamples articles containing French text 10× to address language imbalance.
199
+ * **BE-All:** Combines quality filtering (score ≥ 3), upsampling of clinical content, French text, and clinical cases, plus metadata prefixing.
200
+
201
+ ### Performance Results
202
+
203
+ **SOTA Models for reference**
204
+
205
+ | Model | MedQA | MedMCQA | PubMedQA | Anat | Clin | Bio | Med | Gen | Prof | Avg |
206
+ | ----------------- | :---: | :-----: | :------: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
207
+ | Llama-3-8B | 59.70 | 57.47 | 74.80 | 68.89 | 74.72 | 78.47 | 61.85 | 83.00 | 70.22 | 69.90 |
208
+ | Meditron-70B | 57.10 | 46.80 | 76.60 | 53.30 | 66.70 | 76.30 | 63.00 | 69.00 | 71.60 | 64.49 |
209
+
210
+ **Benchmark Results by Dataset Variant (continual pre-training of OLMo2-7B-stage1)**
211
+
212
+ | Variant | MedQA | MedMCQA | PubMedQA | Anat | Clin | Bio | Med | Gen | Prof | Avg |
213
+ | --------------- | :------------: | :------------: | :------------: | :------------: | :------------: | :------------: | :------------: | :------------: | :------------: | :------------: |
214
+ | OLMo2-7B-stage1 | 45.33 | 41.14 | 75.60 | 54.81 | 63.40 | 69.44 | 53.18 | 69.00 | 59.93 | 59.09 |
215
+ | BE-Base | 44.85 | 41.91 | 76.40 | 57.04 | 64.15 | **70.83** | **59.54** | 69.00 | 59.93 | 60.41 |
216
+ | BE-Clinical | 41.95 | 39.35 | 76.60 | 53.33 | 63.40 | 65.28 | 58.38 | 66.00 | **63.97** | 58.70 |
217
+ | BE-ClinicalCase | 42.11 | 39.52 | 76.60 | 57.04 | 64.91 | 66.67 | **59.54** | 69.00 | 62.87 | 59.81 |
218
+ | BE-Prefix | 45.72 | 41.76 | **77.80** | 57.04 | 64.53 | 68.75 | 57.23 | 66.00 | 61.76 | 60.07 |
219
+ | BE-Educational | 45.64 | **43.08** | 77.00 | 57.04 | 65.28 | 68.06 | 56.65 | **71.00** | 58.82 | 60.29 |
220
+ | BE-All | **47.21** | 42.79 | 76.60 | **60.00** | **65.66** | 68.06 | 58.96 | 69.00 | 61.40 | **61.08** |
221
+
222
+ *Note: The first three columns represent Medical QA benchmarks. The following six (Anat, Clin, Bio, Med, Gen, Prof) are sub-tasks from MMLU Medical. MMLU abbreviations: Anat=Anatomy, Clin=Clinical Knowledge, Bio=College Biology, Med=College Medicine, Gen=Medical Genetics, Prof=Professional Medicine.*
223
+
224
+ ### Results Analysis
225
+
226
+ **Overall performance.** BE-All achieved the highest average performance across benchmarks at 61.08%, surpassing BE-Base (60.41%) by a small but consistent margin (+0.67 pts, see table above). Its strongest improvements appeared in MedQA (47.21%), MMLU Anatomy (60.00%), and Clinical Knowledge (65.66%), suggesting the effectiveness of combining multiple targeted enrichment strategies.
227
+
228
+ **Clinical enrichment.** Clinical enrichment (BE-Clinical) significantly boosted performance on MMLU Professional Medicine benchmark (63.97%, +4.04 pts vs. BE-Base, Figure 2). This improvement was stable from early training, highlighting how clinical narratives enhance the model’s clinical reasoning abilities efficiently.
229
+
230
+ **Educational filtering.** Educational filtering (BE-Educational) consistently improved performance on medical question-answering tasks, notably Medical Genetics (71.00%, +2 pts), MedMCQA (43.08%, +1.17 pts), and PubMedQA (77.00%, +0.6 pts). These tasks likely benefit from the knowledge present in educationally high-quality paragraphs (Figure 2).
231
+
232
+ **Metadata prefixing.** Metadata prefixing (BE-Prefix) specifically improved performance on PubMedQA (77.80%, +1.4 pts vs. BE-Base). Providing explicit paragraph-level metadata helped primarily with structured document comprehension, but it had limited benefits for other tasks.
233
+
234
+ **General biomedical knowledge trade-off.** BE-Base performed better on College Biology (70.83%) than others. Building a biology variant (BE-Bio) could be an interesting future direction, as the current dataset does not specifically target this domain.
235
+
236
+ **Non-English enrichment.** BE-French showed clear improvements in French medical QA (FrenchMedMCQA), achieving 40.5% accuracy, significantly surpassing BE-Base and the OLMo2-7B-stage1 baseline (38.32%, Figure 1). These results illustrate effective adaptation to non-English contexts without modifying the underlying model architecture.
237
+
238
+ ![Benchmark Results](french_medmcqa_performance.png)
239
+
240
+ **Data efficiency and training stability.** As shown in Figure 2, BE-All reached robust benchmark performance using roughly one-third of the tokens required by BE-Base. Individual enrichments (Educational, Clinical) also displayed early and stable improvements, underscoring potential reductions in training time and computational cost.
241
+
242
+ ![Benchmark Results](combined_plot.png)
243
+
244
+ ## Licensing
245
+
246
+ The Biomed-Enriched annotations (document type, domain, educational quality scores, and metadata) are released under the MIT License.
247
+
248
+ The textual content licensing depends on the individual article licenses from PubMed Central Open Access. Each paragraph includes a `license_url` field pointing to the specific license. Users must comply with the respective license terms when using the textual data.
249
+
250
+ ## How to Cite
251
+
252
+ Please cite Biomed-Enriched using:
253
+
254
+ ```
255
+ [coming soon]
256
+ ```