LLM-Assisted Code Cleaning For Training Accurate Code Generators

Authors: Naman Jain, Tianjun Zhang, Wei-Lin Chiang, Joseph E. Gonzalez, Koushik Sen, Ion Stoica

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on two challenging algorithmic code generation benchmarks and find that fine-tuning CODELLAMA-7B on our transformed modularized programs improves the performance by up to 30% compared to fine-tuning on the original dataset.
Researcher Affiliation Academia Naman Jain, Tianjun Zhang, Wei-Lin Chiang, Joseph E. Gonzalez, Koushik Sen & Ion Stoica University of California, Berkeley {naman_jain,tianjunz,weichiang,jegonzal,ksen,istoica}@berkeley.edu
Pseudocode No The paper describes the steps for data cleaning and transformation in text, but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper refers to using CODELLAMA-7B model checkpoint from Hugging Face and VLLM for inference, but does not state that the authors are releasing their own code for the described methodology.
Open Datasets Yes We use two standard algorithmic code generation benchmarks, APPS and CODE-CONTESTS. The benchmarks provide a collection of problem statements described in natural language and corresponding test cases.
Dataset Splits No Table 1: Details about the number of problems, the median number of test cases per problem, and the number of solutions in the APPS and CODE-CONTESTS datasets. (Table only lists 'train' and 'test' counts for problems and solutions, no validation split)
Hardware Specification Yes We train the models for two epochs on the APPS dataset and one epoch on the CODE-CONTESTS dataset using a 5e 5 learning rate and an effective batch size of 256 on 4 A6000 GPUs.
Software Dependencies No The paper mentions using CODELLAMA-7B model, Hugging Face, VLLM, DeepSpeed, and gaoya, but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We train the models for two epochs on the APPS dataset and one epoch on the CODE-CONTESTS dataset using a 5e 5 learning rate and an effective batch size of 256 on 4 A6000 GPUs.