SigmaJahan/deep4ge-dataset
Deep4ge: 14,227 mutation-induced DNN training logs for fault detection and diagnosis research (ICST 2026 Data Showcase)
What's novel
Deep4ge: 14,227 mutation-induced DNN training logs for fault detection and diagnosis research (ICST 2026 Data Showcase)
Code Analysis
17 files read · 4 roundsA dataset of 14,227 DNN training logs (from 59 StackOverflow-derived seed programs) with an AST-based mutation framework that injects faults into FNN/CNN/RNN programs and a Keras callback that logs 31 per-epoch diagnostic features for fault detection research.
Strengths
The AST-based mutation approach is clean and well-organized with a clear base class, 10 fault categories, and a comprehensive 31-column logging callback that captures gradient statistics, activation health, and hardware metrics. The dataset itself (14K training logs with rich per-epoch features) is a genuine contribution to the SE/DL fault detection literature.
Weaknesses
16 of the 34 claimed operators (activation, weight, and training-data categories) use a completely different interface (apply_to_model/apply_to_data) that is incompatible with the AST-based CLI pipeline, making them dead code in the documented workflow. The ChangeLearningRate operator has a critical bug (replaces optimizer with a scalar), the callback silently mutates the learning rate every epoch (interfering with mutations under test), and there are no unit tests for the mutation operators.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
5
Contributors
1
Files
14324
Active weeks
2
Repository
Language
Python
Stars
1
Forks
0
License
NOASSERTION