🚀 JumpForge-3K

Agentic Coding Traces for Modern Software Engineering

جامپ‌فورج-۳کی | مجموعه داده‌ای برای عامل‌های کدنویس و مهندسی نرم‌افزار مدرن

---

Overview

JumpForge-3K is a synthetic dataset designed for training and evaluating agentic coding systems.

The dataset focuses on realistic software engineering workflows, including repository understanding, debugging, code review, test-driven development, tool-use planning, and multi-file reasoning.

Built by JumpLander, this dataset serves as a foundation for experimentation in agentic AI and coding assistants.

Key Capabilities

• Repository understanding
• Multi-file reasoning
• Bug fixing workflows
• Code review tasks
• Test-first development
• Refactoring scenarios
• API design reasoning
• Tool-use planning

---

Dataset Information

Property	Value
Dataset Name	JumpForge-3K
Total Samples	3,000
Train Split	2,700
Validation Split	300
Languages	English, Persian
License	CC BY 4.0
Format	JSONL

---

Task Categories

Task Type	Description
Repository Understanding	Analyze project structure and architecture
Bug Fixing	Identify and resolve software defects
Code Review	Review and improve existing code
Test-Driven Development	Design implementations from tests
Refactoring	Improve maintainability and structure
API Design	Create or improve API interfaces
Debugging Traces	Reason through failure scenarios
Tool Use Planning	Plan coding-agent tool execution

---

Dataset Schema

Each JSONL record contains the following fields:

Field	Type	Description
id	string	Unique sample identifier
dataset	string	Dataset name
brand	string	Branding information
task_type	string	Task category
language	string	Programming language
difficulty	string	Difficulty level
title	string	Task title
domain	string	Domain or industry
user_story	string	User scenario
repo_context	string	Repository context
instruction	string	Agent instruction
reference_solution	string	Reference solution outline
evaluation	object	Evaluation criteria
tags	array	Metadata tags

---

Example Record

{
  "id": "JT3K-000722",
  "dataset": "JumpForge-3K",
  "brand": "JumpForge-3K",
  "task_type": "tool_use",
  "language": "python",
  "difficulty": "easy",
  "title": "Sequence tool calls",
  "domain": "creator-tools",
  "user_story": "Construct a tool-use plan for fixing an import error across multiple files.",
  "repo_context": "Repo: ml-pipeline. Domain: creator-tools.",
  "instruction": "Create a tool-use plan and define the optimal workflow.",
  "reference_solution": "Inspect files, isolate failure source, patch minimally, run tests, verify results.",
  "evaluation": {
    "checks": [
      "tool order sensible",
      "minimal edit strategy",
      "verification included"
    ]
  },
  "tags": [
    "agentic",
    "coding",
    "tool_use",
    "python"
  ]
}

---

Loading the Dataset

Install Dependencies

pip install datasets

Load Dataset

from datasets import load_dataset

dataset = load_dataset("jumplander/JumpForge-3K")

train_dataset = dataset["train"]
validation_dataset = dataset["validation"]

print(train_dataset[0]["title"])

---

Training Example (Unsloth)

from datasets import load_dataset
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/Qwen2.5-Coder-1.5B-Instruct",
    load_in_4bit=True,
)

dataset = load_dataset(
    "jumplander/JumpForge-3K",
    split="train"
)

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

### Repository Context:
{example['repo_context']}

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

dataset = dataset.map(format_example)

---

Suggested Use Cases

Use Case	Description
SFT Training	Fine-tuning coding assistants
Agent Research	Agent workflow experiments
Repository Reasoning	Multi-file reasoning benchmarks
Tool Use Studies	Tool-planning evaluation
DPO/RLHF Prototyping	Preference optimization research
Persian Coding Research	Persian-focused coding experiments

---

Dataset Statistics

Languages

Language	Approximate Count
English	2,500+
Persian	500+

Difficulty Distribution

Difficulty	Approximate Count
Easy	1,000
Medium	1,500
Hard	500

Programming Languages

Language	Approximate Count
Python	2,000+
JavaScript	500+
SQL	300+
TypeScript	150+
Bash	50+

---

Notes

This dataset is synthetic and intended primarily for:

• Format prototyping
• Agent workflow experimentation
• Benchmark development
• Fine-tuning research

It should not be considered sufficient as the sole training source for production-grade coding models.

---

License

This dataset is distributed under the CC BY 4.0 License.

Attribution is required when redistributing or creating derivative works.

---

Citation

@misc{jumpforge3k,
  title={JumpForge-3K: Agentic Coding Traces for Modern Software Engineering},
  author={JumpLander Team},
  year={2026},
  publisher={Hugging Face}
}

---

🚀 Built by JumpLander

Official Website: jumplander.org
Hugging Face Dataset: JumpForge-3K

---

About the Brand: This dataset and the associated "Trace" format were created by JumpLander. The organization focuses on Agentic Coding Traces and modern software engineering workflows. The logo is designed to represent "JumpLander" and is associated with the official domain and Hugging Face organization page.

---