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 meta-llama/Llama-2-7b-chat-hf and has been fine-tuned for two epochs.

It achieves the following results on the evaluation set:

Loss: 0.8173

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: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.0345	0.1988	123	0.9800
0.8537	0.3976	246	0.8808
0.5807	0.5964	369	0.8503
0.7419	0.7952	492	0.8413
0.9967	0.9939	615	0.8406
0.7252	1.1939	738	0.8301
0.9605	1.3927	861	0.8214
0.7785	1.5915	984	0.8186
0.7233	1.7903	1107	0.8178
0.8389	1.9891	1230	0.8173

Training time: it took 80 minutes to fine-tune the model for two epochs.

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/AmsterdamDocClassificationLlama200T2Epochs

AmsterdamDocClassificationLlama200T2Epochs