# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("jrc/phi3-mini-math", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("jrc/phi3-mini-math", trust_remote_code=True)

Training Details

Phi3 was trained using torchtune and the training script + config file are located in this repository.

tune run lora_finetune_distributed.py --config mini_lora.yaml

You can see a full Weights & Biases run here.

Training Data

This model was finetuned on the following datasets:

TIGER-Lab/MATH-plus: An advanced math-specific dataset with 894k samples.

Hardware

Machines: 4 x NVIDIA A100 GPUs
Max VRAM used per GPU: 29 GB
Real time: 10 hours

Evaluation

The finetuned model is evaluated on minerva-math using EleutherAI Eval Harness through torchtune.

bash

tune run eleuther_eval --config eleuther_evaluation \
          checkpoint.checkpoint_dir=./lora-phi3-math \
          tasks=["minerva_math"] \
          batch_size=32 

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
minerva_math	N/A	none	4	exact_match	0.1670	±	0.0051
- minerva_math_algebra	1	none	4	exact_match	0.2502	±	0.0126
- minerva_math_counting_and_prob	1	none	4	exact_match	0.1329	±	0.0156
- minerva_math_geometry	1	none	4	exact_match	0.1232	±	0.0150
- minerva_math_intermediate_algebra	1	none	4	exact_match	0.0576	±	0.0078
- minerva_math_num_theory	1	none	4	exact_match	0.1148	±	0.0137
- minerva_math_prealgebra	1	none	4	exact_match	0.3077	±	0.0156
- minerva_math_precalc	1	none	4	exact_match	0.0623	±	0.0104

This shows a large improvement over the base Phi3 Mini model.

Model Card Contact

Drop me a line at @official_j3rck