README.md

---
license: apache-2.0
pipeline_tag: question-answering
tags:
- question-answering
- transformers
- generated_from_trainer
datasets:
- squad_v2
- LLukas22/nq-simplified
- newsqa
- LLukas22/NLQuAD
- deepset/germanquad
language:
- en
- de
---

# all-MiniLM-L12-v2-qa-all
This model is an extractive qa model.
It's a fine-tuned version of [all-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2) on the following datasets: [squad_v2](https://huggingface.co/datasets/squad_v2), [LLukas22/nq-simplified](https://huggingface.co/datasets/LLukas22/nq-simplified), [newsqa](https://huggingface.co/datasets/newsqa), [LLukas22/NLQuAD](https://huggingface.co/datasets/LLukas22/NLQuAD), [deepset/germanquad](https://huggingface.co/datasets/deepset/germanquad).



## Usage

You can use the model like this:

```python
from transformers import pipeline

#Make predictions
model_name = "LLukas22/all-MiniLM-L12-v2-qa-all"
nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)

QA_input = {
    "question": "What's my name?",
    "context": "My name is Clara and I live in Berkeley."
}

result = nlp(QA_input)
print(result)
```
Alternatively you can load the model and tokenizer on their own:
```python
from transformers import AutoModelForQuestionAnswering, AutoTokenizer

#Make predictions
model_name = "LLukas22/all-MiniLM-L12-v2-qa-all"
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
```

## Training hyperparameters
The following hyperparameters were used during training:

- learning_rate: 2E-05
- per device batch size: 60
- effective batch size: 180
- seed: 42
- optimizer: AdamW with betas (0.9,0.999) and eps 1E-08
- weight decay: 1E-02
- D-Adaptation: False
- Warmup: True
- number of epochs: 15
- mixed_precision_training: bf16

## Training results
| Epoch | Train Loss | Validation Loss |
| ----- | ---------- | --------------- |
| 0 | 3.76 | 3.02 |
| 1 | 2.57 | 2.23 |
| 2 | 2.2 | 2.08 |
| 3 | 2.07 | 2.03 |
| 4 | 1.96 | 1.97 |
| 5 | 1.87 | 1.93 |
| 6 | 1.81 | 1.91 |
| 7 | 1.77 | 1.89 |
| 8 | 1.73 | 1.89 |
| 9 | 1.7 | 1.9 |
| 10 | 1.68 | 1.9 |
| 11 | 1.67 | 1.9 |

## Evaluation results
| Epoch | f1 | exact_match |
| ----- | ----- | ----- |
| 0 | 0.29 | 0.228 |
| 1 | 0.371 | 0.329 |
| 2 | 0.413 | 0.369 |
| 3 | 0.437 | 0.376 |
| 4 | 0.454 | 0.388 |
| 5 | 0.468 | 0.4 |
| 6 | 0.479 | 0.408 |
| 7 | 0.487 | 0.415 |
| 8 | 0.495 | 0.421 |
| 9 | 0.501 | 0.416 |
| 10 | 0.506 | 0.42 |
| 11 | 0.51 | 0.421 |

## Framework versions
- Transformers: 4.25.1
- PyTorch: 2.0.0.dev20230210+cu118
- PyTorch Lightning: 1.8.6
- Datasets: 2.7.1
- Tokenizers: 0.13.1
- Sentence Transformers: 2.2.2

## Additional Information
This model was trained as part of my Master's Thesis **'Evaluation of transformer based language models for use in service information systems'**. The source code is available on [Github](https://github.com/LLukas22/Retrieval-Augmented-QA).