arnavgrg/codealpaca-qlora

QLoRA weights using Llama-2-7b for the Code Alpaca Dataset

Fine-Tuning on Predibase

This model was fine-tuned using Predibase, the first low-code AI platform for engineers. I fine-tuned base Llama-2-7b using LoRA with 4 bit quantization on a single T4 GPU, which cost approximately $3 to train on Predibase. Try out our free Predibase trial here.

Dataset and training parameters are borrowed from: https://github.com/sahil280114/codealpaca, but all of these parameters including DeepSpeed can be directly used with Ludwig, the open-source toolkit for LLMs that Predibase is built on.

Co-trained by: Infernaught

How To Use The Model

To use these weights:

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM

# Load base model in 4 bit
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", load_in_4bit=True)

# Wrap model with pretrained model weights
config = PeftConfig.from_pretrained("arnavgrg/codealpaca-qlora")
model = PeftModel.from_pretrained(model, "arnavgrg/codealpaca-qlora")

Prompt Template:

Below is an instruction that describes a task, paired with an input
that provides further context. Write a response that appropriately
completes the request.

### Instruction: {instruction}

### Input: {input}

### Response:

Training procedure

The following bitsandbytes quantization config was used during training:

• —loadin8bit: False
• —loadin4bit: True
• —llmint8threshold: 6.0
• —llmint8skip_modules: None
• —llmint8enablefp32cpu_offload: False
• —llmint8hasfp16weight: False
• —bnb4bitquant_type: nf4
• —bnb4bitusedoublequant: True
• —bnb4bitcompute_dtype: float16

Framework versions

• —PEFT 0.4.0