andre156/italian-laws-topic-extraction

Model Description

A Mistral-7B-instruct-v0.1 model to extract the topics from a text of Italian law articles (titles). It is fine-tuned over a set of 74k high quality law title-topics pairs, which were initially obtained from the application of a larger model (Mixtral8x22) and then pre-processed to increase the quality of the training set by means of heuristics that aggregate slighlty different topics and allow the fine-tuned model to achieve an higher diversity

• Developed by: Andrea Colombo, Politecnico di Milano
• Model type: text generation
• Language(s) (NLP): Italian
• License: Apache 2.0
• Finetuned from model: mistralai/Mistral-7B-Instruct-v0.1

How to Get Started with the Model

Training Details

Training Procedure

The model has been trained for 100 training steps with batch size 4, 4-bit quantization using bitsandbytes and a LoRA rank of 64. We use the paged Adam optimizer, a learning rate of 0.004, and a cosine learning rate scheduler with a 0.03 warm-up fraction.

Evaluation

The best model reported an evaluation loss of 0.61