magistrate-3.2-3b-it

This model is a fine-tuned version of macadeliccc/magistrate-3.2-3b-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8067

👁 Built with Axolotl

Model description

Magistrate-3.2-3b-it is a legal assistant specializing in US Supreme Court case law and US Federal regulations.

The base model is pretrained with ~250M tokens containing no synthetic legal data. The instruct model does contain synthetic data.

Intended uses & limitations

This model is for research purposes and for continued development of the legal specialty. You are liable for all model outputs.

Training and evaluation data

This model was trained on a variety of standard open source datasets like OpenHermes-2.5, hermes-function-calling, and some select entries from the Tome. Additionally, I have included a comprehensive, non-synthetic argument dataset. This is a work in progress but has shown promising results so far.

Training procedure

Spectrum top 35% finetune for both pretrain and SFT. Thanks to the cognitive computations team for the work done with spectrum.

• Pretraining methodology based on Cohere's paper: To Code, or Not To Code? Exploring Impact of Code in Pre-training
• Instruct finetune largely based on OpenHermes-2.5 and hermes-function-calling

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• total_eval_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 1000
• num_epochs: 3

Training results

Framework versions

• Transformers 4.45.0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.20.0

Base model

meta-llama/Llama-3.2-3B

macadeliccc/magistrate-3.2-3b-base

(1)