DDIDU/ETRI_CodeLLaMA_7B_CPP

ETRI_CodeLLaMA_7B_CPP

We used LoRa to further pre-train Meta's CodeLLaMA-7B-hf model with high-quality C++ code tokens.

Furthermore, we fine-tuned on CodeM's C++ instruction data.

Model Details

This model was trained using LoRa and achieved a pass@1 of 34.3% on HumanEvalX-cpp.

ETRI_CodeLLaMA_7B_CPP is a C++ specialized model.

Dataset Details

We pre-trained CodeLLaMA-7B further using 543 GB of C++ code collected online, and fine-tuned it using CodeM's C++ instruction data. We utilized 1 x A100-80GB GPU for the training.

Requirements

pip install torch transformers accelerate

How to reproduce HumanEval-X results

We use Bigcode-evaluation-harness repo for evaluating our trained model.

bigcode-evaluation-harness

git clone https://github.com/bigcode-project/bigcode-evaluation-harness.git

Then, run main.py as follows.

accelerate launch bigcode-evaluation-harness/main.py \
  --model DDIDU/ETRI_CodeLLaMA_7B_CPP \
  --max_length_generation 512 \
  --prompt continue \
  --tasks humanevalsynthesize-cpp \
  --temperature 0.2 \
  --n_samples 100 \
  --precision bf16 \
  --do_sample True \
  --batch_size 10 \
  --allow_code_execution \
  --save_generations \

Model use

from transformers import AutoTokenizer
import transformers
import torch

model = "DDIDU/ETRI_CodeLLaMA_7B_CPP"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

sequences = pipeline(
    '#include <iostream>\n#include <vector>\n\nusing namespace std;\n\nvoid quickSort(int *data, int start, int end) {',
    do_sample=True,
    top_k=10,
    temperature=0.1,
    top_p=0.95,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    max_length=200,
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")