Datasets:

DanCip
/

lca-StartingPoints-expanded

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - code
  - python
arxiv: 2508.02455
dataset_info:
  features:
    - name: idx
      dtype: int64
    - name: idx_lca
      dtype: int64
    - name: offset
      dtype: int64
    - name: repo
      dtype: string
    - name: commit_hash
      dtype: string
    - name: target_file
      dtype: string
    - name: line_type_lca
      dtype: string
    - name: ground_truth
      dtype: string
    - name: in_completions
      dtype: bool
    - name: completion_type
      dtype: string
    - name: non_dunder_count_intellij
      dtype: int64
    - name: non_dunder_count_jedi
      dtype: int64
    - name: start_with_
      dtype: bool
    - name: first_occurrence
      dtype: bool
    - name: intellij_completions
      sequence: string
    - name: jedi_completions
      list:
        - name: name
          dtype: string
        - name: type
          dtype: string
    - name: prefix
      dtype: string
  splits:
    - name: train
      num_bytes: 76152400
      num_examples: 5531
  download_size: 8547476
  dataset_size: 76152400
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

🧠 LCA-Starting Points

A benchmark for evaluating project-local code completion ranking. Curated to validate TreeRanker (ASE2025).

📖 Dataset Description

Starting Points is a specialized dataset designed to evaluate code completion ranking, with a specific focus on locally defined identifiers (project-specific APIs) in Python.

Most LLM benchmarks focus on global APIs (standard libraries). However, developers spend significant time using APIs defined within their own projects. Starting Points targets this "blind spot" by testing how well models can resolve and rank identifiers that are defined within the user's current repository but may not be visible in the immediate file context.

This dataset is a refined subset of the Long Code Arena, enriched with:

Static Analysis Data: Valid completions resolved by the Jedi library.
Real-World IDE Suggestions: Ranked candidate lists generated by IntelliJ IDEA.

⚡ Key Features

Focus: Project-specific API completion (vs. standard library).
Language: Python.
Source: Large projects with rich user-defined classes and functions.
Goal: Benchmark ranking algorithms for local development environments.

📂 Dataset Structure

Data Instances

Each instance represents a specific cursor position in a Python file where a dereference operation (e.g., object.) occurs. The task is to predict the correct next identifier from a list of candidates.

📊 Data Fields

This dataset includes rich metadata to facilitate deep analysis of ranking performance, surpassing the original details in the paper.

Field	Type	Description
`idx`	`int64`	Unique identifier for the dataset entry.
`idx_lca`	`int64`	Original index of the file in the Long Code Arena benchmark.
`repo`	`string`	Name of the source repository.
`commit_hash`	`string`	Specific commit hash used for the snapshot.
`target_file`	`string`	Path to the file within the repository.
`offset`	`int64`	Character offset (cursor position) where completion is triggered.
`prefix`	`string`	Source code content preceding the cursor (the context).
`ground_truth`	`string`	The actual identifier the developer typed (target label).
`completion_type`	`string`	Metadata describing the completion scenario.
`start_with_`	`bool`	True if the ground truth starts with an underscore `_`.
`first_occurrence`	`bool`	True if the identifier has not appeared previously in the `prefix` (file context). Useful for evaluating "unseen" identifier performance.
`in_completions`	`bool`	True if `ground_truth` is present in the `intellij_completions` list.
`intellij_completions`	`sequence[string]`	Ranked list of candidates from IntelliJ IDEA's completion engine.
`jedi_completions`	`list[struct]`	List of valid completions from Jedi, including `name` and `type`.
`non_dunder_count_intellij`	`int64`	Count of IntelliJ candidates excluding dunder methods (e.g., `__init__`).
`non_dunder_count_jedi`	`int64`	Count of Jedi candidates excluding dunder methods.
`line_type_lca`	`string`	Inherited metadata from Long Code Arena regarding line classification.

🛠️ Dataset Creation

Curation Rationale

Standard benchmarks often overlook the difficulty of ranking identifiers that are local to a specific project. This dataset was created to test model performance in this realistic, everyday scenario where context from other files in the repository is crucial.

Source Data

Origin: Long Code Arena benchmark.
Filtering:
- Dereference Detection: Uses tree-sitter to find . operators.
- Static Analysis: Uses Jedi to resolve object types.
- Scope Constraint: Retains only suggestions defined within the same repository, excluding standard library calls.
- Quality Control: Requires the ground truth to be in the IntelliJ candidate list and the list to have at least 5 non-trivial items.

📚 Citation

If you use this dataset, please cite the original paper:

@article{cipollone2025treeranker,
  title={TreeRanker: Fast and Model-agnostic Ranking System for Code Suggestions in IDEs},
  author={Cipollone, Daniele and Bogomolov, Egor and van Deursen, Arie and Izadi, Maliheh},
  journal={arXiv preprint arXiv:2508.02455},
  year={2025}
}