jumplander/J6-CFI-HQ-20K

J6-CFI-HQ-20K: JumpLander High-Quality 20K Code Feedback Instructions

A high-quality 20K English code instruction dataset for coding assistants, debugging, code review, refactoring, test generation, software engineering reasoning, and coding-agent behavior.

Built by [JumpLander](https://www.jumplander.org) for experiments in coding assistants, software engineering datasets, and agentic developer intelligence.

<p align="center"> <img src="https://huggingface.co/datasets/jumplander/J6-CFI/resolve/main/J6-CFI.webp" alt="J6-CFI Dataset Overview - JumpLander Code Feedback Instructions" width="100%" /> </p>

Overview

J6-CFI-HQ-20K stands for JumpLander High-Quality 20K Code Feedback Instructions.

This dataset is an English-only, synthetic, high-quality structured code-instruction dataset designed for supervised fine-tuning experiments, coding assistant prototyping, debugging behavior, code review, refactoring, test generation, algorithmic reasoning, backend/API task design, SQL and data engineering tasks, and coding-agent behavior research.

Unlike minimal instruction -> answer datasets, J6-CFI-HQ-20K includes structured metadata for each record: developer intent, constraints, reference solution, explanation, test cases, quality signals, evaluation checks, and curation metadata.

This release is intentionally smaller than large synthetic corpora. The focus is cleaner structure, stronger constraints, more realistic developer tasks, and better record-level metadata.

Dataset Information

Task Distribution

Programming Language Distribution

Quality Positioning

J6-CFI-HQ-20K is not a human-verified Gold benchmark. It is a high-quality synthetic dataset candidate designed to be cleaner and more focused than large-scale raw synthetic corpora.

Generation and curation properties:

• —20,000 valid JSONL records
• —exact duplicate full records: 0
• —exact duplicate instructions: 0
• —exact duplicate code contexts: 0
• —each record includes explicit constraints
• —each record includes test cases or validation scenarios
• —each record includes quality signals and evaluation checks
• —each record includes curation metadata
• —designed for SFT and coding-assistant prototyping

For stricter evaluation, a future J6-CFI-Gold-1K subset should be manually reviewed.

Dataset Schema

Each JSONL line contains one record with the following fields:

Loading the Dataset

pip install datasets

from datasets import load_dataset

dataset = load_dataset("jumplander/J6-CFI-HQ-20K")
train = dataset["train"]

print(train[0]["instruction"])
print(train[0]["reference_solution"])

Load Locally

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={{"train": "j6_cfi_hq_20k.jsonl"}}
)

Formatting for SFT

def format_example(example):
    return {{
        "text": f"""### Instruction:
{{example['instruction']}}

### Code Context:
{{example['code_context']}}

### Developer Intent:
{{example['intent']}}

### Constraints:
{{example['constraints']}}

### Reference Solution:
{{example['reference_solution']}}

### Explanation:
{{example['explanation']}}

### Verification:
{{example['test_cases']}}"""
    }}

formatted = dataset["train"].map(format_example)

Suggested Use Cases

Important Quality Notes

J6-CFI-HQ-20K is a high-quality synthetic structured dataset, not a fully human-verified benchmark.

It is suitable for SFT experiments, coding assistant prototyping, structured instruction tuning, behavior-aware code generation research, and preparing a manually reviewed Gold subset.

It should not be described as human-verified, production-certified, a complete correctness benchmark, or a replacement for execution-based evaluation.

Recommended next steps:

1. 1.create J6-CFI-Gold-1K,
2. 2.manually review technical correctness,
3. 3.execute a subset of Python, JavaScript, SQL, and Bash examples,
4. 4.score solution usefulness and difficulty,
5. 5.publish an evaluation script.

Dataset Philosophy

A strong coding assistant should not only produce code. It should understand developer intent, preserve constraints, avoid unsafe rewrites, explain trade-offs, include tests, and produce actionable engineering output.

J6-CFI-HQ-20K is designed around that principle.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Attribution is required when redistributing, adapting, or building upon this dataset.

Citation

@misc{{j6_cfi_hq_20k_2026,
  title={{J6-CFI-HQ-20K: JumpLander High-Quality 20K Code Feedback Instructions}},
  author={{JumpLander Team}},
  year={{2026}},
  publisher={{Hugging Face}},
  url={{https://huggingface.co/datasets/jumplander/J6-CFI-HQ-20K}}
}}

Built by JumpLander

J6-CFI-HQ-20K is built by [JumpLander](https://www.jumplander.org), a research and engineering initiative focused on coding agents, developer intelligence, AI-assisted programming, and practical software engineering datasets.

• —Website: jumplander.org
• —Persian Website: jumplander.org/fa
• —Documentation: jumplander.org/fa/docs
• —About: jumplander.org/fa/about
• —Hugging Face: huggingface.co/jumplander