README.md

---
language:
- 'no'
- nb
- nn
license: cc-by-4.0
datasets:
- ltg/norec_sentence
pipeline_tag: text-classification
---

# Sentence-level Sentiment Analysis model for Norwegian text
This model is a fine-tuned version of [ltg/norbert3-base](https://huggingface.co/ltg/norbert3-base) for text classification.

## Training data
The dataset used for fine-tuning is [ltg/norec_sentence](https://huggingface.co/datasets/ltg/norec_sentence), the `mixed` subset with four sentement categories:  
```
[0]:  Negative,  
[1]:  Positive,  
[2]:  Neutral
[0,1]: Mixed
```

## Quick start
You can use this model for inference as follows:
```
>>> from transformers import pipeline
>>> origin = "ltg/norbert3-large_sentence-sentiment"
>>> pipe = transformers.pipeline( "text-classification",
...                             model = origin,
...                             trust_remote_code=origin.startswith("ltg/norbert3"),
...                             config= origin,
...                             tokenizer = AutoTokenizer.from_pretrained(origin)
...             )
>>> preds = pipe(["Hans hese, litt såre stemme kler bluesen, men denne platen kommer neppe til å bli blant hans største kommersielle suksesser.",
...              "Borten-regjeringen gjorde ikke jobben sin." ])
>>> for p in preds:
...     print(p)
```
Output:
```
The model 'NorbertForSequenceClassification' is not supported for text-classification. Supported models are ['AlbertForSequenceClassification', ...
{'label': 'Mixed', 'score': 0.7435498237609863}
{'label': 'Negative', 'score': 0.765734851360321}
```

## Training hyperparameters
- per_device_train_batch_size: 32
- learning_rate: 1e-05
- gradient_accumulation_steps: 1
- num_train_epochs: 10 (best epoch 2)

## Evaluation
| Category          |       F1 | |
|:----------------|---------:|----:|
| Negative_F1     | 0.670241 |<img width=400/> |
| Positive_F1     | 0.832918 | |
| Neutral_F1      | 0.850082 | |
| Mixed_F1        | 0.580645 | |
| Weighted_avg_F1 | 0.799663 | |