phi-2-basic-maths

This model is a fine-tuned version of microsoft/phi-2 on an GSM8K dataset.

Model Description

The objective of this model is to evaluate Phi-2's ability to provide correct solutions to reasoning problems after fine-tuning. This model was trained using techniques such as TRL, LoRA quantization, and Flash Attention.

To test it, you can use the following code:

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer, pipeline

# Specify the model ID
peft_model_id = "Menouar/phi-2-basic-maths"

# Load Model with PEFT adapter
model = AutoPeftModelForCausalLM.from_pretrained(
  peft_model_id,
  device_map="auto",
  torch_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained(peft_model_id)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

Training procedure

The complete training procedure can be found on my Notebook.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 42
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 84
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 30

Training results

The training results can be found on Tensoboard.

Evaluation procedure

The complete Evaluation procedure can be found on my Notebook.

Accuracy: 36.16%

Unclear answers: 7.81%

Framework versions

• PEFT 0.8.2
• Transformers 4.38.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	53.60
AI2 Reasoning Challenge (25-Shot)	55.80
HellaSwag (10-Shot)	71.15
MMLU (5-Shot)	47.27
TruthfulQA (0-shot)	41.40
Winogrande (5-shot)	75.30
GSM8k (5-shot)	30.71