naimsassine/mistralinstruct-7b-sft-lora-belgianlaw

mistralinstruct-7b-sft-lora

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on the Belgian Law QnA dataset. It achieves the following results on the evaluation set:

• Loss: 1.2937

Model description

The goal of this model was to experiment how far we can push a model by fine tuning it on a french QnA dataset on Belgian Law. The goal here is to see if we can get a small size LLM to become good enough in terms of Legal Expertise in a specific country

Intended uses & limitations

Legal Question Answering (Belgian Law/French)

Training and evaluation data

SFT-LORA Big thanks to Niels Rogge's notebook that helped me through the process https://github.com/NielsRogge/Transformers-Tutorials/blob/master/Mistral/Supervised_fine_tuning_(SFT)_of_an_LLM_using_Hugging_Face_tooling.ipynb

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 128
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.3021	0.9778	10	1.2937

Framework versions

• PEFT 0.12.0
• Transformers 4.42.4
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1