oceann315/mistral-7b-qlora-arc_challenge-epoch3-v1.2.2

Model Card: Mistral 7B LoRA Fine-tuned on ARC Challenge

Model Details

• Model Name: oceann315/mistral-7b-qlora-arc_challenge-epoch3-v1.2.2
• Base Model: mistralai/Mistral-7B-v0.1
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Number of Trainable Parameters: 46,137,344 / 3,798,208,512
• Dataset: ARC Challenge
• Objective: Improving performance on ARC Challenge, a dataset for multiple-choice question answering

Training Details

• Hardware: Google Colab A100
• Batch Size: 64
• Learning Rate: 2e-4
• LoRA Rank: 32
• LoRA Alpha: 64
• Number of Training Steps/Epochs: 3
• Optimizer: paged_adamw_32bit
• Scheduler: constant

Model Performance

• Evaluation Metric: acc, acc_norm
• Baseline Score (Mistral 7B v0.1): 0.5700, 0.6126
• Fine-tuned Model Score: 0.5546, 0.6007

Limitations and Considerations

This model is fine-tuned for the ARC Challenge dataset and may not generalize well to other tasks.
It may still struggle with certain types of reasoning questions.
The model inherits biases from both the base model and the dataset.
Further improvements could be explored by incorporating additional datasets, refining hyperparameters, or experimenting with different LoRA configurations.

Usage

• Intended Use: Answering multiple-choice science questions
• How to Use:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "oceann315/mistral-7b-qlora-arc_challenge-epoch3-v1.2.2"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input_text = "Your ARC Challenge question here"
inputs = tokenizer(input_text, return_tensors="pt")
output = model.generate(**inputs)
print(tokenizer.decode(output[0], skip_special_tokens=True))