Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
datasets:
|
4 |
+
- wikipedia
|
5 |
+
language:
|
6 |
+
- en
|
7 |
+
metrics:
|
8 |
+
- glue
|
9 |
+
---
|
10 |
+
# Model Card for SzegedAI/bert-medium-mlsm
|
11 |
+
|
12 |
+
<!-- Provide a quick summary of what the model is/does. -->
|
13 |
+
|
14 |
+
This medium-sized BERT model was created using the [Masked Latent Semantic Modeling] (MLSM) pre-training objective, which is a sample efficient alternative for classic Masked Language Modeling (MLM).
|
15 |
+
During MLSM, the objective is to recover the latent semantic profile of the masked tokens, as opposed to recovering their exact identity.
|
16 |
+
The contextualized latent semantic profile during pre-training is determined by performing sparse coding of the hidden representation of an already pre-trained model (a base-sized BERT model in this particular case).
|
17 |
+
|
18 |
+
## Model Details
|
19 |
+
|
20 |
+
### Model Description
|
21 |
+
|
22 |
+
<!-- Provide a longer summary of what this model is. -->
|
23 |
+
|
24 |
+
- **Developed by:** SzegedAI
|
25 |
+
- **Model type:** transformer encoder
|
26 |
+
- **Language:** Engish
|
27 |
+
- **License:** MIT
|
28 |
+
|
29 |
+
### Model Sources
|
30 |
+
|
31 |
+
<!-- Provide the basic links for the model. -->
|
32 |
+
|
33 |
+
- **Repository:** [https://github.com/szegedai/MLSM](https://github.com/szegedai/MLSM)
|
34 |
+
- **Paper:** [Masked Latent Semantic Modeling: an Efficient Pre-training Alternative to Masked Language Modeling](https://underline.io/events/395/posters/15279/poster/78046-masked-latent-semantic-modeling-an-efficient-pre-training-alternative-to-masked-language-modeling?tab=abstract+%26+voting)
|
35 |
+
|
36 |
+
## How to Get Started with the Model
|
37 |
+
|
38 |
+
The pre-trained model can be used in the usual manner, e.g., for fine tuning on a particular sequence classification task, invoke the code:
|
39 |
+
|
40 |
+
```
|
41 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
42 |
+
|
43 |
+
tokenizer = AutoTokenizer.from_pretrained('SzegedAI/bert-medium-mlsm')
|
44 |
+
model = AutoModelForSequenceClassification.from_pretrained('SzegedAI/bert-medium-mlsm')
|
45 |
+
```
|
46 |
+
|
47 |
+
## Training Details
|
48 |
+
|
49 |
+
### Training Data
|
50 |
+
|
51 |
+
<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
|
52 |
+
|
53 |
+
The model was pre-trained using a 2022 English Wikipedia dump pre-processed with [wiki-bert-pipeline](https://github.com/spyysalo/wiki-bert-pipeline).
|
54 |
+
|
55 |
+
### Training Procedure
|
56 |
+
|
57 |
+
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
|
58 |
+
|
59 |
+
#### Preprocessing
|
60 |
+
|
61 |
+
|
62 |
+
#### Training Hyperparameters
|
63 |
+
|
64 |
+
Pre-training was conducted with a batch size of 32 sequences and a gradient accumulation over 32 batches, resulting in an effective batch size of 1024.
|
65 |
+
A total of 300,000 update steps were performed using the AdamW optimizer with a linear learning rate scheduling having a peak learning rate of 1e-04.
|
66 |
+
A maximum sequence length of 128 tokens was employed over the first 90% of the pre-training, while for the final 10% of the pre-training, the maximum sequence length was increased to 512 tokens.
|
67 |
+
|
68 |
+
- **Training regime:** fp32
|
69 |
+
|
70 |
+
## Evaluation
|
71 |
+
|
72 |
+
<!-- This section describes the evaluation protocols and provides the results. -->
|
73 |
+
|
74 |
+
#### Metrics
|
75 |
+
|
76 |
+
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
|
77 |
+
|
78 |
+
The model was evaluated on GLUE tasks and CoNLL2003 for named entity recognition.
|
79 |
+
|
80 |
+
### Results
|
81 |
+
The evaluation result after fine-tuning the given model on a wide range of tasks.
|
82 |
+
On each tasks 10 different fine-tuning were performed, during which the only difference was the random initialization of the task-specific classification head.
|
83 |
+
Both the average and the standard deviation are displayed below on each tasks.
|
84 |
+
|
85 |
+
| Dataset | Metric | Avg. | Std. |
|
86 |
+
|---|---|---|---|
|
87 |
+
| CoLA | Matthews correlation | 0.403 | 0.012 |
|
88 |
+
| CoNLL2003 | F1 | 0.926 | 0.003 |
|
89 |
+
| MNLI (matched) | Accuracy | 0.798 | 0.001 |
|
90 |
+
| MNLI (mismatched) | Accuracy | 0.808 | 0.002 |
|
91 |
+
| MRPC | Accuracy | 0.786 | 0.020 |
|
92 |
+
| MRPC | F1 | 0.851 | 0.013 |
|
93 |
+
| QNLI | Accuracy | 0.870 | 0.004 |
|
94 |
+
| QQP | Accuracy | 0.892 | 0.001 |
|
95 |
+
| QQP | F1 | 0.855 | 0.001 |
|
96 |
+
| RTE | Accuracy | 0.571 | 0.011 |
|
97 |
+
| SST2 | Accuracy | 0.905 | 0.004 |
|
98 |
+
| STSB | Pearson correlation | 0.818 | 0.024 |
|
99 |
+
| STSB | Spearman correlation | 0.820 | 0.021 |
|
100 |
+
| WiC | Accuracy | 0.639 | 0.007 |
|
101 |
+
| Average | --- | 0.7815 | --- |
|
102 |
+
#### Summary
|
103 |
+
|
104 |
+
This model was more sample efficient and reached practically the same average performance as an alternatively pre-trained language model of 2.5 times more parameter (of base size)
|
105 |
+
that was pre-trained using the classical MLM objective.
|
106 |
+
|
107 |
+
## Environmental Impact
|
108 |
+
|
109 |
+
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
110 |
+
|
111 |
+
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
|
112 |
+
|
113 |
+
- **Hardware Type:** RTX A6000
|
114 |
+
- **Hours used:** 300
|
115 |
+
- **Carbon Emitted:** 42 kg CO2 eq.
|
116 |
+
|
117 |
+
## Citation
|
118 |
+
|
119 |
+
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
120 |
+
The pre-training objective is introduced in the ACL Findings paper _Masked Latent Semantic Modeling: an Efficient Pre-training Alternative to Masked Language Modeling_.
|
121 |
+
|
122 |
+
**BibTeX:**
|
123 |
+
|
124 |
+
|
125 |
+
|