LLMcoder - A Feedback-Based Coding Assistant

Practical AI Methods and Tools for Programming

Introduction

Enhancemenent of coding assistants through integration of feedback from LLM querying. LLMCoder fetches and retrieves API information and documentation based on error report obtained from four different analyzers. These aim to prevent instances of type errors, incorrect API signature usage and LLM-induced hallucinations. These have been implemented after fine-tuning GPT-3.5 aligning with a evaluated scale of difficulty levels.

Requirements

Hardware

Software

• Python >= 3.10
• pip >= 21.3

  Getting Started

  Clone The Repository

git clone https://github.com/pvs-hd-tea/23ws-LLMcoder
cd 23ws-LLMcoder

Create A Virtual Environment (optional):

With conda

conda create -n llmcoder python=3.11 [ipykernel]
conda activate llmcoder

Install

Install the package

pip install -e .

Usage

CLI:

llmcoder [command] [options]

Python API:

from llmcoder import LLMcoder

llmcoder = LLMcoder(
    analyzers=[
        "mypy_analyzer_v1",  # Detect and fix type errors
        "signature_analyzer_v1",  # Augment type errors with signatures
        "gpt_score_analyzer_v1"],  # Score and find best completion
    feedback_variant="coworker",  # Make context available to all analyzers
    max_iter=3,  # Maximum number of feedback iterations
    n_procs=4  # Complete and analyze in parallel
    backtracking=True, # Enable backtracking in the tree of completions
    verbose=True  # Print progress
)

# Define an incomplete code snippet
code = "print("

# Complete the code
result = llmcoder.complete(code, n=4)

Evaluation

1. Compile the dataset

To compile a dataset from input-output-pairs to a conversations.jsonl file, run

llmcoder export -n name/of/dataset

on a dataset stored in /data/name/of/dataset.

2. Install packages used in the evaluation

pip install -r data/name/of/dataset/requirements.txt

3. Run the LLMcoder Evaluation

To evaluate LLMcoder on all configs in /configs, run

llmcoder evaluate

To evaluate LLMcoder on a specific config, run

llmcoder evaluate -c my_config.yaml

where my_config.yaml is a configuration file from /configs.

The following files will be created for each config and run:

• /data/name/of/dataset/eval/<config_name>/<run_id>/results.json with
  • messages: the message history
  • analyzer_results: analyzer results for each step
  • log: a log of the run
  • time: the time it took to complete the run
• /data/name/of/dataset/eval/<config_name>/<run_id>/readable_logs/<example_id>.txt with
  • a human-readable log of the run

for each example in the dataset

4. Compute metrics

After running the evaluation, compute the metrics for all configs in /configs with

llmcoder metrics

To compute the metrics for a specific config, run

llmcoder metrics -c my_config.yaml

where my_config.yaml is a configuration file from /configs.#

This will create the following files for each config and run:

• /data/name/of/dataset/eval/<config_name>/<run_id>/metrics.csv

Development

Setup

To set up the development environment, run the following commands:

pip install -e .[dev]
pre-commit install

For further information, see CONTRIBUTING.md.

Citation

@software{llmcoder-hd-24,
    author = {Ana Carsi and Kushal Gaywala and Paul Saegert},
    title = {LLMcoder: Feedback-Based Code Assistant},
    month = mar,
    year = 2024,
    publisher = {GitHub},
    version = {0.4.0},
    url = {https://github.com/pvs-hd-tea/23ws-LLMcoder}
}