As part of the Assessing Large Language Models for Document Classification project by the Municipality of Amsterdam, we fine-tune Mistral, Llama, and GEITje for document classification. The fine-tuning is performed using the AmsterdamBalancedFirst200Tokens dataset, which consists of documents truncated to the first 200 tokens. In our research, we evaluate the fine-tuning of these LLMs across one, two, and three epochs. This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.2 and has been fine-tuned for one epoch.

It achieves the following results on the evaluation set:

Loss: 0.7673

Training and evaluation data

The training data consists of 9900 documents and their labels formatted into conversations.
The evaluation data consists of 1100 documents and their labels formatted into conversations.

Training procedure

See the GitHub for specifics about the training and the code.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 2
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.7202	0.1988	123	0.8994
0.7523	0.3976	246	0.8242
0.4706	0.5964	369	0.7938
0.5285	0.7952	492	0.7687
0.7677	0.9939	615	0.7673

Training time: it took in total 44 minutes to fine-tuned the model for one epoch.

Framework versions

Transformers 4.41.1
Pytorch 2.3.0+cu121
Datasets 2.19.1
Tokenizers 0.19.1

Acknowledgements

This model was trained as part of [insert thesis info] in collaboration with Amsterdam Intelligence for the City of Amsterdam.

FemkeBakker/AmsterdamDocClassificationMistral200T1Epochs

AmsterdamDocClassificationMistral200T1Epochs