README.md

metadata

license: apache-2.0
datasets:
  - verifiers-for-code/CodeNet-16K
  - verifiers-for-code/CodeNet-Planner
language:
  - en
library_name: transformers
tags:
  - code

🦙 Llama-3-LlamaPlanner

Model Description

LlamaPlanner is a fine-tuned version of Meta's Llama-8B model which has been specifically designed for generating high-quality plans for code generation tasks. The model was trained on CodeNet-16k, a curated dataset of competitive programming problems, and their corresponding plans generated using Llama-3-70B. By leveraging the power of Parameter Efficient Fine-Tuning (PEFT), LlamaPlanner achieves performance comparable to much larger models in generating effective plans for code generation.

Model Details

• Base Model: Llama-8B Instruct
• Fine-Tuning Approach: Parameter Efficient Fine-Tuning (PEFT) using Unsloth
• Training Data: CodeNet-16k, a filtered and deduplicated dataset of 16,500 competitive programming problems and their plans generated using Llama-3-70B
• Training Infrastructure: H100-SXM5 GPU
• Evaluation Benchmarks: HumanEval and EvalPlus

How to Use

To use LlamaPlanner with the Hugging Face Transformers library, follow these steps:

import transformers
import torch

model_id = "verifiers-for-code/Llama-3-LlamaPlanner"

pipeline = transformers.pipeline(
    "text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto"
)

prompt = "Generate a plan for a program that sorts an array of integers in ascending order."
pipeline(prompt)

Training Details

LlamaPlanner was trained using the following steps:

1. Filtering and preprocessing the CodeNet dataset to create CodeNet-16k
2. Generating plans for each problem using Llama-3-70B
3. Formatting the problem description, input description, output description, and samples as input, and the generated plans as output
4. Performing PEFT on the Llama-8B Instruct base model using Unsloth with different ranks and alpha values
5. Training on an H100-SXM5 GPU for varying epochs

Evaluation Results

LlamaPlanner was evaluated on the HumanEval and EvalPlus benchmarks using various methods, including zero-shot, self-planning, base planner model, and fine-tuned planner model. The results demonstrated that LlamaPlanner outperforms the base Llama-3-8B model by 14% on HumanEval and 11% on EvalPlus. Additionally, plans generated by LlamaPlanner helped boost the performance of Llama-3-70B on HumanEval.

Citation

If you use LlamaPlanner in your research or applications, please cite the model using the following BibTeX entry:

@misc{llamaplanner,
  title={LlamaPlanner: A Fine-Tuned Llama-8B Model for Effective Plan Generation in Code Generation Tasks},
  author={Abhinav Chinta and Sumuk Shashidhar and Vaibhav Sahai},
  year={2023},
  howpublished={\url{https://huggingface.co/verifiers-for-code/LlamaPlanner}},
}

License

LlamaPlanner is released under the Apache License 2.0.

Acknowledgements

We would like to thank Meta for releasing the Llama model family and the open-source community for their contributions to the development of large language models and their applications in code generation tasks.