Hugging Face's logo

MikaSie
/
RoBERTa_T5_dependent_V1

Summarization
Transformers
Safetensors
English
t5
text2text-generation
abstractive
hybrid
multistep
Eval Results
text-generation-inference
Inference Endpoints
Model card Files Community
MikaSie commited on
Commit
e6894bd
•
1 Parent(s): 12ff612

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +105 -80
README.md CHANGED
@@ -1,85 +1,110 @@
1
  ---
2
- license: apache-2.0
3
- base_model: t5-large
4
  tags:
5
- - generated_from_trainer
6
- datasets:
7
- - arrow
 
 
 
 
8
  model-index:
9
- - name: RoBERTa_T5_dependent_V1
10
- results: []
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
  ---
12
 
13
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
14
- should probably proofread and complete it, then remove this comment. -->
15
-
16
- # RoBERTa_T5_dependent_V1
17
-
18
- This model is a fine-tuned version of [t5-large](https://huggingface.co/t5-large) on the arrow dataset.
19
- It achieves the following results on the evaluation set:
20
- - Loss: 2.4943
21
-
22
- ## Model description
23
-
24
- More information needed
25
-
26
- ## Intended uses & limitations
27
-
28
- More information needed
29
-
30
- ## Training and evaluation data
31
-
32
- More information needed
33
-
34
- ## Training procedure
35
-
36
- ### Training hyperparameters
37
-
38
- The following hyperparameters were used during training:
39
- - learning_rate: 5e-05
40
- - train_batch_size: 1
41
- - eval_batch_size: 1
42
- - seed: 42
43
- - distributed_type: multi-GPU
44
- - num_devices: 4
45
- - gradient_accumulation_steps: 4
46
- - total_train_batch_size: 16
47
- - total_eval_batch_size: 4
48
- - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
49
- - lr_scheduler_type: linear
50
- - lr_scheduler_warmup_ratio: 0.1
51
- - num_epochs: 40
52
-
53
- ### Training results
54
-
55
- | Training Loss | Epoch | Step | Validation Loss |
56
- |:-------------:|:-------:|:----:|:---------------:|
57
- | 9.2486 | 0.9963 | 68 | 4.2390 |
58
- | 3.8105 | 1.9927 | 136 | 3.2877 |
59
- | 3.2106 | 2.9890 | 204 | 2.9341 |
60
- | 2.8659 | 4.0 | 273 | 2.7676 |
61
- | 2.7264 | 4.9963 | 341 | 2.6706 |
62
- | 2.5997 | 5.9927 | 409 | 2.6175 |
63
- | 2.5016 | 6.9890 | 477 | 2.5781 |
64
- | 2.3939 | 8.0 | 546 | 2.5520 |
65
- | 2.3659 | 8.9963 | 614 | 2.5352 |
66
- | 2.3132 | 9.9927 | 682 | 2.5209 |
67
- | 2.2625 | 10.9890 | 750 | 2.5084 |
68
- | 2.1871 | 12.0 | 819 | 2.5006 |
69
- | 2.1777 | 12.9963 | 887 | 2.4977 |
70
- | 2.1434 | 13.9927 | 955 | 2.4919 |
71
- | 2.1108 | 14.9890 | 1023 | 2.4913 |
72
- | 2.0496 | 16.0 | 1092 | 2.4878 |
73
- | 2.0483 | 16.9963 | 1160 | 2.4907 |
74
- | 2.0225 | 17.9927 | 1228 | 2.4924 |
75
- | 1.9965 | 18.9890 | 1296 | 2.4926 |
76
- | 1.9427 | 20.0 | 1365 | 2.4960 |
77
- | 1.9492 | 20.9963 | 1433 | 2.4943 |
78
-
79
-
80
- ### Framework versions
81
-
82
- - Transformers 4.40.1
83
- - Pytorch 2.2.1+cu121
84
- - Datasets 2.17.1
85
- - Tokenizers 0.19.1
 
 
 
 
1
  ---
2
+ language: en
 
3
  tags:
4
+ - summarization
5
+ - abstractive
6
+ - hybrid
7
+ - multistep
8
+ datasets: dennlinger/eur-lex-sum
9
+ pipeline_tag: summarization
10
+ base_model: T5
11
  model-index:
12
+ - name: BART
13
+ results:
14
+ - task:
15
+ type: summarization
16
+ name: Long, Legal Document Summarization
17
+ dataset:
18
+ name: eur-lex-sum
19
+ type: dennlinger/eur-lex-sum
20
+ metrics:
21
+ - type: ROUGE-1
22
+ value: 0.2933917923410164
23
+ - type: ROUGE-2
24
+ value: 0.09261541731427447
25
+ - type: ROUGE-L
26
+ value: 0.18570590575800888
27
+ - type: BERTScore
28
+ value: 0.8404191312843194
29
+ - type: BARTScore
30
+ value: -1.5701540725191212
31
+ - type: BLANC
32
+ value: 0.1112198320851479
33
  ---
34
 
35
+ # Model Card for RoBERTa_T5_dependent_V1
36
+
37
+ ## Model Details
38
+ ---
39
+ ### Model Description
40
+
41
+ This model is a fine-tuned version of T5. The research involves a multi-step summarization approach to long, legal documents. Many decisions in the renewables energy space are heavily dependent on regulations. But these regulations are often long and complicated. The proposed architecture first uses one or more extractive summarization steps to compress the source text, before the final summary is created by the abstractive summarization model. This fine-tuned abstractive model has been trained on a dataset, pre-processed through extractive summarization by RoBERTa with dependent ratio. The research has used multiple extractive-abstractive model combinations, which can be found on https://huggingface.co/MikaSie. To obtain optimal results, feed the model an extractive summary as input as it was designed this way!
42
+
43
+ The dataset used by this model is the [EUR-lex-sum](https://huggingface.co/datasets/dennlinger/eur-lex-sum) dataset. The evaluation metrics can be found in the metadata of this model card.
44
+ This paper was introduced by the master thesis of Mika Sie at the University Utrecht in collaboration with Power2x. More information can be found in PAPER_LINK.
45
+
46
+ - **Developed by:** Mika Sie
47
+ - **Funded by:** University Utrecht & Power2X
48
+ - **Language (NLP):** English
49
+ - **Finetuned from model:** T5
50
+
51
+
52
+ ### Model Sources
53
+
54
+ - **Repository**: https://github.com/MikaSie/Thesis
55
+ - **Paper**: PAPER_LINK
56
+ - **Streamlit demo**: STREAMLIT_LINK
57
+
58
+ ## Uses
59
+ ---
60
+ ### Direct Use
61
+
62
+ This model can be directly used for summarizing long, legal documents. However, it is recommended to first use an extractive summarization tool, such as RoBERTa, to compress the source text before feeding it to this model. This model has been specifically designed to work with extractive summaries.
63
+ An example using the Huggingface pipeline could be:
64
+
65
+ ```python
66
+ pip install bert-extractive-summarizer
67
+
68
+ from summarizer import Summarizer
69
+ from transformers import pipeline
70
+
71
+ extractive_model = Summarizer()
72
+
73
+ text = 'Original document text to be summarized'
74
+
75
+ extractive_summary = Summarizer(text)
76
+
77
+ abstractive_model = pipeline('summarization', model = 'MikaSie/RoBERTa_T5_dependent_V1', tokenizer = 'MikaSie/RoBERTa_T5_dependent_V1')
78
+
79
+ result = pipeline(extractive_summary)
80
+ ```
81
+
82
+ But more information of implementation can be found in the Thesis report.
83
+ ### Out-of-Scope Use
84
+
85
+ Using this model without an extractive summarization step may not yield optimal results. It is recommended to follow the proposed multi-step summarization approach outlined in the model description for best performance.
86
+
87
+ ## Bias, Risks, and Limitations
88
+ ---
89
+
90
+ ### Bias
91
+
92
+ As with any language model, this model may inherit biases present in the training data. It is important to be aware of potential biases in the source text and to critically evaluate the generated summaries.
93
+
94
+ ### Risks
95
+
96
+ - The model may not always generate accurate or comprehensive summaries, especially for complex legal documents.
97
+ - The model may not generate truthful information.
98
+
99
+ ### Limitations
100
+
101
+ - The model may produce summaries that are overly abstractive or fail to capture important details.
102
+ - The model's performance may vary depending on the quality and relevance of the extractive summaries used as input.
103
+
104
+ ### Recommendations
105
+
106
+ - Carefully review and validate the generated summaries before relying on them for critical tasks.
107
+ - Consider using the model in conjunction with human review or other validation mechanisms to ensure the accuracy and completeness of the summaries.
108
+ - Experiment with different extractive summarization models or techniques to find the most suitable input for the abstractive model.
109
+ - Provide feedback and contribute to the ongoing research and development of the model to help improve its performance and address its limitations.
110
+ - Any actions taken based on this content are at your own risk.