File size: 9,607 Bytes
e2260ee e37091f e2260ee e37091f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 |
---
language:
- en
license: apache-2.0
tags:
- biomedical
- clinical
- ul2
- t5
- encoder-decoder
- pretraining
- text2text-generation
- medical
---
# PubMedUL2 & MedUL2
## Model Description
**PubMedUL2** and **MedUL2** are a family of **domain-specific UL2/T5-style encoder–decoder language models** pretrained on large-scale biomedical and medical corpora using the **UL2 (Mixture-of-Denoisers)** objective.
- **PubMedUL2** models are pretrained on **25 million PubMed abstracts**
- **MedUL2** models are pretrained on **PubMed abstracts + clinical notes + additional medical documents**
- All models use a **T5-efficient architecture**, inspired by Google’s efficient T5 variants
These checkpoints are **pretraining-only models** and **must be fine-tuned** before use on downstream tasks.
---
## Pretraining Objective: UL2 (Mixture-of-Denoisers)
These models were pretrained using **UL2**, a unified framework that formulates language modeling objectives as **denoising tasks**.
UL2 introduces a **Mixture-of-Denoisers (MoD)** approach that samples from multiple denoising paradigms during pretraining.
### Denoising Tasks
UL2 pretraining uses a mixture of three denoising tasks:
1. **R-denoising (Regular Span Corruption)**
- Equivalent to standard T5 span corruption
- Optimized for language understanding tasks
2. **X-denoising (Extreme Span Corruption)**
- Uses very large masked spans
- Encourages long-form generation and abstraction
3. **S-denoising (Sequential / PrefixLM)**
- Prefix language modeling similar to causal LM
- Suitable for sequence-to-sequence and generative tasks
### Paradigm Tokens (Mode Switching)
During pretraining, a **paradigm token** is inserted at the beginning of each input:
| Token | Mode | Recommended Use |
|------|------|------------------|
| `[NLU]` | R-denoising | Classification, QA, retrieval |
| `[NLG]` | X-denoising | Mixed understanding & generation |
| `[S2S]` | S-denoising | Generative / causal tasks |
**Important:**
For best performance, the same token should be **prepended during fine-tuning and inference**.
---
## Architecture
- Encoder–decoder Transformer (T5-style)
- Uses **T5-efficient architecture**
- Compatible with Hugging Face `T5ForConditionalGeneration`
---
## Intended Uses
These models are intended to be **fine-tuned** for:
- Biomedical and clinical **text classification**
- **Question answering**
- **Summarization** of medical literature or clinical notes
- **Text generation** in medical contexts
---
## Limitations
- ❌ Not instruction-tuned
- ❌ No supervised training
- ❌ Not suitable for zero-shot use
These checkpoints are **self-supervised pretraining models only** and require task-specific fine-tuning.
---
## Fine-Tuning Recommendations
- **Avoid mixed precision** (fp16 / bf16) initially
- Fine-tuning is more stable in **fp32**
- Always prepend one of `[NLU]`, `[NLG]`, or `[S2S]` to input text
- Suggested defaults:
- Classification / QA → `[NLU]`
- Causal or generative tasks → `[S2S]`
- Mixed tasks → `[NLG]`
---
## Model Parameter Summary
| Model Name | Parameter Count | Description | Access
|-----------|----------------|------------|------------|
| `pubmedul2-tiny-nl6` | **19.26M** | Tiny UL2-style model with 6 layers | Open
| `pubmedul2-mini-nl8` | **50.12M** | Mini UL2 with 8 layers | Open
| `pubmedul2-small` | **60.52M** | Small UL2 variant | Open
| `pubmedul2-small-nl24` | **192.73M** | Small UL2 with 24 layers | Open
| `medul2-base` | **222.93M** | Base UL2/T5-style model | Open
| `pubmedul2-base` | **222.93M** | Base UL2/T5-style model | Open
| `medul2-base-nl36` | **619.44M** | Base UL2 with 36 layers | Gated commercial
| `pubmedul2-base-nl36` | **619.44M** | Base UL2 with 36 layers | Gated commercial
| `medul2-large` | **737.72M** | Large UL2/T5-style model | Gated non-commercial
| `pubmedul2-large` | **737.72M** | Large UL2/T5-style model | Gated non-commercial
| `medul2-large-nl36` | **1090.14M** | Very large UL2 with 36 layers | Access on Request
---
## Named Entity Recognition (NER) Evaluation
We evaluate PubMedUL2 and MedUL2 models on a biomedical **Named Entity Recognition (NER)** task using multiple matching criteria to better capture boundary-level performance.
The evaluation reports **entity-level F1 scores** across different biomedical entity types and model sizes.
### Exact Match F1
An entity prediction is considered correct only if both the **entity span and label exactly match** the gold annotation.
| entity_type | medul2-base | pubmedul2-base | pubmedul2-mini-nl8 | pubmedul2-small | pubmedul2-tiny-nl6 |
|:--------------|--------------:|-----------------:|---------------------:|------------------:|---------------------:|
| cell_line | 0.42 | 0.43 | 0.44 | 0.43 | 0.35 |
| cell_type | 0.59 | 0.58 | 0.59 | 0.58 | 0.52 |
| chemical | 0.76 | 0.75 | 0.72 | 0.72 | 0.56 |
| disease | 0.7 | 0.73 | 0.7 | 0.68 | 0.63 |
| dna | 0.59 | 0.55 | 0.54 | 0.55 | 0.45 |
| gene | 0.62 | 0.59 | 0.6 | 0.59 | 0.55 |
| protein | 0.59 | 0.58 | 0.58 | 0.59 | 0.55 |
| rna | 0.6 | 0.56 | 0.55 | 0.6 | 0.56 |
| species | 0.66 | 0.67 | 0.58 | 0.63 | 0.54 |
---
### Partial Match F1
A prediction is counted as correct if it **partially overlaps** with a gold entity of the same type.
| entity_type | medul2-base | pubmedul2-base | pubmedul2-mini-nl8 | pubmedul2-small | pubmedul2-tiny-nl6 |
|:--------------|--------------:|-----------------:|---------------------:|------------------:|---------------------:|
| cell_line | 0.48 | 0.49 | 0.48 | 0.48 | 0.41 |
| cell_type | 0.66 | 0.64 | 0.66 | 0.65 | 0.59 |
| chemical | 0.79 | 0.78 | 0.76 | 0.75 | 0.6 |
| disease | 0.82 | 0.84 | 0.8 | 0.79 | 0.74 |
| dna | 0.65 | 0.61 | 0.6 | 0.61 | 0.53 |
| gene | 0.76 | 0.74 | 0.74 | 0.73 | 0.68 |
| protein | 0.66 | 0.66 | 0.66 | 0.67 | 0.64 |
| rna | 0.68 | 0.63 | 0.64 | 0.66 | 0.65 |
| species | 0.68 | 0.7 | 0.61 | 0.65 | 0.56 |
---
### IoU Match F1
Predictions are evaluated using **Intersection-over-Union (IoU)** overlap between predicted and gold spans, providing a softer boundary-based metric.
| entity_type | medul2-base | pubmedul2-base | pubmedul2-mini-nl8 | pubmedul2-small | pubmedul2-tiny-nl6 |
|:--------------|--------------:|-----------------:|---------------------:|------------------:|---------------------:|
| cell_line | 0.5 | 0.5 | 0.5 | 0.5 | 0.42 |
| cell_type | 0.67 | 0.66 | 0.68 | 0.67 | 0.62 |
| chemical | 0.83 | 0.83 | 0.82 | 0.82 | 0.72 |
| disease | 0.85 | 0.86 | 0.86 | 0.85 | 0.82 |
| dna | 0.65 | 0.62 | 0.62 | 0.62 | 0.55 |
| gene | 0.76 | 0.75 | 0.75 | 0.74 | 0.71 |
| protein | 0.67 | 0.66 | 0.67 | 0.67 | 0.66 |
| rna | 0.68 | 0.65 | 0.66 | 0.67 | 0.67 |
| species | 0.72 | 0.74 | 0.65 | 0.69 | 0.58 |
---
### Observations
- **MedUL2 models** generally outperform PubMedUL2 on clinical-heavy entity types such as *disease* and *chemical*
- Performance improves consistently from **tiny → base models**
- Boundary-sensitive metrics (Partial / IoU) show significantly higher scores than Exact Match, highlighting boundary ambiguity in biomedical NER
---
## Acknowledgements
This project would not have been possible without compute generously provided by **Google TPU Research Cloud**.
Thanks to:
- The **Finnish-NLP** authors for releasing the UL2 objective code, task definitions, and guidance
- **Yeb Havinga** for help getting started with the **t5x** framework
---
## License
Please refer to the individual model repositories for **license and access details**, which may vary depending on training data sources.
|