JumpForge-Agentic-SE-3K

JumpForge-Agentic-SE-3K is a structured synthetic dataset for training and evaluating AI software-engineering agents. Its primary target is agent behavior across the software engineering lifecycle, not raw code generation or memorization of programming-language syntax.

The dataset teaches an agent to:

• understand intent and ambiguity before acting;
• explore repositories and trace system behavior;
• decompose work into reversible steps;
• select tools based on information value;
• debug causally instead of patching symptoms;
• preserve contracts, security, and operational safety;
• verify changes with failure-focused tests;
• recover from failed actions;
• calibrate uncertainty and escalate responsibly;
• communicate evidence, risk, and remaining unknowns;
• maintain state across long-horizon engineering tasks.

Dataset size

Split	Records	Archetype groups
Train	2400	96
Validation	300	12
Test	300	12
Total	3,000	120

Each archetype contains 25 independently parameterized scenarios. Archetypes are kept entirely inside one split to reduce behavioral-template leakage between train, validation, and test.

Behavioral taxonomy

The dataset has 12 top-level families and 120 archetypes:

1. Problem understanding
2. Repository exploration
3. Planning and decomposition
4. Tool selection
5. Debugging behavior
6. Safe modification
7. Verification and testing
8. Error recovery
9. Security-aware engineering
10. Uncertainty and escalation
11. Communication and reporting
12. Long-horizon execution

See docs/TAXONOMY.md and configs/taxonomy.json.

Record design

Each record includes:

• user request and system context;
• repository structure and relevant files;
• evidence, logs, and a focused failing test;
• hidden ground truth for evaluation;
• ideal agent assessment, assumptions, plan, and tool trace;
• observations, decision, implementation strategy, verification, and stopping criteria;
• bad-behavior counterexamples;
• evaluation rubric with must-do and must-not-do conditions;
• chat-style messages for SFT-compatible workflows;
• machine-validation metadata and hashes.

Example access

from datasets import load_dataset

dataset = load_dataset("jumplander/JumpForge-Agentic-SE-3K")
print(dataset["train"][0]["task"])

Local JSONL:

import json

with open("data/train.jsonl", encoding="utf-8") as f:
    first = json.loads(next(f))

print(first["ideal_agent_behavior"]["plan"])

Deduplication and leakage controls

The release includes machine-readable reports under reports/.

Current generated-release audit:

• exact duplicate IDs: 0
• exact duplicate user requests: 0
• normalized duplicate user requests: 0
• duplicate final responses: 0
• duplicate structural fingerprints: 0
• archetype split leakage: 0
• maximum TF-IDF nearest-neighbor similarity: 0.7350

The lexical check is a heuristic. It should not be interpreted as proof that all records are semantically independent. Before high-stakes model training or research publication, perform expert review and embedding/model-assisted semantic audits.

Important limitations

• The dataset is synthetic.
• Machine validation checks structure, consistency constraints, uniqueness, and split policy.
• The included heuristic quality score is not human quality certification.
• Code paths and evidence are realistic abstractions, not executable full repositories.
• Tool traces describe expected agent decisions; they are not logs from tools actually executed.
• Some phrasing and structural regularity remains because the dataset is generated from a controlled taxonomy.

Recommended uses

• supervised fine-tuning for agent planning and communication;
• behavior scoring and evaluator development;
• tool-selection and stopping-criteria research;
• repository-reasoning curriculum design;
• preference-data generation using the included bad behaviors;
• benchmark prototyping for safe software-engineering agents.

Not recommended

• measuring real-world patch correctness without executable repositories;
• claiming production-grade coding performance from this dataset alone;
• security certification;
• direct autonomous execution on sensitive systems.

License

CC BY 4.0. See LICENSE.

Attribution

Created by JumpLander for research on agentic software engineering.