README.md

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- code
- python
- long-context
- coding
size_categories:
- 1K<n<10K
configs:
- config_name: 0k
  data_files:
  - split: test
    path: data/0k/test/data-*
  - split: train
    path: data/0k/train/data-*
  - split: validation
    path: data/0k/validation/data-*
  - split: prompt
    path: data/0k/prompt/data-*
- config_name: 1k
  data_files:
  - split: test
    path: data/1k/test/data-*
  - split: train
    path: data/1k/train/data-*
  - split: validation
    path: data/1k/validation/data-*
  - split: prompt
    path: data/1k/prompt/data-*
- config_name: 2k
  data_files:
  - split: test
    path: data/2k/test/data-*
  - split: train
    path: data/2k/train/data-*
  - split: validation
    path: data/2k/validation/data-*
  - split: prompt
    path: data/2k/prompt/data-*
- config_name: 4k
  data_files:
  - split: test
    path: data/4k/test/data-*
  - split: train
    path: data/4k/train/data-*
  - split: validation
    path: data/4k/validation/data-*
  - split: prompt
    path: data/4k/prompt/data-*
- config_name: 8k
  data_files:
  - split: test
    path: data/8k/test/data-*
  - split: train
    path: data/8k/train/data-*
  - split: validation
    path: data/8k/validation/data-*
  - split: prompt
    path: data/8k/prompt/data-*
- config_name: 16k
  data_files:
  - split: test
    path: data/16k/test/data-*
  - split: train
    path: data/16k/train/data-*
  - split: validation
    path: data/16k/validation/data-*
  - split: prompt
    path: data/16k/prompt/data-*
- config_name: 32k
  data_files:
  - split: test
    path: data/32k/test/data-*
  - split: train
    path: data/32k/train/data-*
  - split: validation
    path: data/32k/validation/data-*
  - split: prompt
    path: data/32k/prompt/data-*
- config_name: 64k
  data_files:
  - split: test
    path: data/64k/test/data-*
  - split: train
    path: data/64k/train/data-*
  - split: validation
    path: data/64k/validation/data-*
  - split: prompt
    path: data/64k/prompt/data-*
- config_name: 128k
  data_files:
  - split: test
    path: data/128k/test/data-*
  - split: train
    path: data/128k/train/data-*
  - split: validation
    path: data/128k/validation/data-*
  - split: prompt
    path: data/128k/prompt/data-*
- config_name: 196k
  data_files:
  - split: test
    path: data/196k/test/data-*
  - split: train
    path: data/196k/train/data-*
  - split: validation
    path: data/196k/validation/data-*
  - split: prompt
    path: data/196k/prompt/data-*
- config_name: 256k
  data_files:
  - split: test
    path: data/256k/test/data-*
  - split: train
    path: data/256k/train/data-*
  - split: validation
    path: data/256k/validation/data-*
  - split: prompt
    path: data/256k/prompt/data-*
- config_name: 512k
  data_files:
  - split: test
    path: data/512k/test/data-*
  - split: train
    path: data/512k/train/data-*
  - split: validation
    path: data/512k/validation/data-*
  - split: prompt
    path: data/512k/prompt/data-*
- config_name: 1m
  data_files:
  - split: test
    path: data/1m/test/data-*
  - split: train
    path: data/1m/train/data-*
  - split: validation
    path: data/1m/validation/data-*
  - split: prompt
    path: data/1m/prompt/data-*
dataset_info:
  features:
  - name: task_id
    dtype: int64
  - name: text
    dtype: string
  - name: code
    dtype: string
  - name: test_list
    sequence: string
  - name: test_setup_code
    dtype: string
  - name: challenge_test_list
    sequence: string
  - name: context
    dtype: string
  - name: context_id
    dtype: string
  - name: context_length_tokens
    dtype: int64
  - name: code_length_chars
    dtype: int64
  - name: dataset_version
    dtype: string
  splits:
  - name: test
    num_examples: 500
  - name: train
    num_examples: 374
  - name: validation
    num_examples: 90
  - name: prompt
    num_examples: 10
---

# MBPP Long-Context Dataset

## Overview

MBPP Long-Context is a benchmark dataset that combines coding problems from the [MBPP (Mostly Basic Python Problems)](https://github.com/google-research/google-research/tree/master/mbpp) dataset with long-context distractors from [BABILong](https://github.com/booydar/babilong). This dataset evaluates code generation performance under long-context conditions, testing whether models can maintain coding ability with stuffed context.

## Dataset Structure

### Data Fields

Each sample contains:

#### Original MBPP Fields
- `task_id` (int): Unique task identifier
- `text` (str): Problem description
- `code` (str): Reference solution
- `test_list` (List[str]): Test cases (assertions)
- `test_setup_code` (str): Optional setup code
- `challenge_test_list` (List[str]): Additional test cases

#### Long-Context Fields
- `context` (str): Prepended distractor text from BABILong, ranging from 0k to 1M.
- `context_id` (str): BABILong source identifier (e.g., "babilong_128k_qa1_sample_42")
- `context_length_tokens` (int): Token count using Llama tokenizer

#### Metadata
- `code_length_chars` (int): Reference solution length for difficulty tracking

### Data Splits

All configurations follow the original MBPP split structure:

- **test**: 500 samples (primary evaluation set)
- **train**: 374 samples
- **validation**: 90 samples
- **prompt**: 10 samples (few-shot examples)

## Creating the dataset

To avoid confounding variables, this dataset uses stratified random assignment, where:

1. Sort MBPP tasks by code length
2. Get text from BABILong qa1-qa10 splits
3. Duplicate contexts to match task count (974 samples)
4. Shuffle contexts and assign to sorted tasks

## Source Datasets

### MBPP (Mostly Basic Python Problems)
- **Source**: [google-research-datasets/mbpp](https://huggingface.co/datasets/google-research-datasets/mbpp)
- **Size**: 974 problems
- **Paper**: [Program Synthesis with Large Language Models](https://arxiv.org/abs/2108.07732)

### BABILong
- **Source**: [RMT-team/babilong](https://huggingface.co/datasets/RMT-team/babilong)
- **Content**: `input` field from qa1-qa10 splits
- **Paper**: [BABILong: Testing the Limits of LLMs with Long Context Reasoning-in-a-Haystack](https://arxiv.org/abs/2406.10149)