Datasets:
Code Debugging Trajectory Dataset - Data Dictionary
===================================================
Generated: 2026-06-26
Sessions: 50,000 | Events: 665,364 | Avg events/session: 13.3
SESSION-LEVEL DATA (code_debugging_sessions.csv)
38 columns describing the overall debugging session
| Column | Type | Description |
|---|---|---|
| session_id | string | Unique identifier (DBG_0000000 format) |
| timestamp_start | datetime | Session start time (2025 calendar year) |
| timestamp_end | datetime | Session end time |
| day_of_week | categorical | Monday-Sunday |
| hour_of_day | int (0-23) | Start hour with realistic work-hour bias |
| is_weekend | bool | True if Saturday or Sunday |
| years_experience | float | Developer years of experience (0-25) |
| experience_level | categorical | Junior/Mid-level/Senior/Staff-Principal |
| editor_used | categorical | IDE/editor (VS Code dominant at 35%) |
| operating_system | categorical | macOS/Linux/Windows |
| company_size | categorical | Startup to Enterprise |
| team_context | categorical | Solo, pair, sprint, on-call, review, learning |
| programming_language | categorical | 8 languages with realistic market share |
| project_type | categorical | Web, mobile, ML, CLI, microservice, etc. |
| codebase_size_loc | int | Lines of code (log-normal, median ~22k) |
| files_modified | int | Number of files touched |
| primary_file_extension | categorical | File extension of primary language |
| error_type | categorical | 71 language-specific error types |
| error_severity | int (1-5) | Severity rating based on error taxonomy |
| error_message_length | int | Character count of error message |
| stack_trace_depth | int | Number of frames in stack trace |
| similar_bug_before | bool | Whether developer encountered similar bug |
| resolution_time_seconds | int | Total debugging time (15s - 2hr cap) |
| resolution_time_minutes | float | Same as above in minutes |
| compile_attempts | int | Number of compile/build attempts |
| num_web_searches | int | Web searches performed |
| external_resources_used | string | Pipe-separated list of resources |
| ai_assistant_used | bool | Whether ChatGPT/Copilot/etc was used |
| asked_colleague | bool | Whether help was requested |
| num_file_navigations | int | File/cursor navigation events |
| lines_changed | int | Lines of code modified |
| keystrokes_per_minute | int | Typing intensity proxy |
| took_break | bool | Break taken during session |
| fix_strategy | categorical | What approach ultimately worked |
| outcome | categorical | fixed / workaround_applied / escalated / abandoned |
| fix_quality | categorical | poor / acceptable / good / excellent / refactored / N/A (N/A when outcome is not 'fixed') |
| regression_introduced | bool | Whether fix caused new bugs |
| test_added | bool | Whether test was added with fix |
EVENT-LEVEL DATA (code_debugging_events.csv)
665,364 individual events across all sessions
| Column | Type | Description |
|---|---|---|
| session_id | string | Foreign key to sessions table |
| event_sequence | int | Order of event within session (0-indexed) |
| event_type | categorical | 15 event types (compile, edit, search, etc.) |
| event_detail | string | Context-specific detail for the event. 'general' where no specific detail applies. |
| event_timestamp | datetime | When event occurred |
| event_duration_seconds | int | How long the event lasted |
| elapsed_seconds | int | Cumulative time since session start |
| session_stage | categorical | early / middle / late |
| is_final_event | bool | Whether this is the last event in session |
KEY DESIGN DECISIONS
Realistic distributions: Language shares match StackOverflow survey data. Python (30%), JavaScript (22%), Java (15%), TypeScript (12%), C++ (8%), Go (5%), C# (4%), Rust (4%).
Experience-dependent behavior: Juniors search more, use AI more, take longer to resolve. Seniors navigate files more efficiently.
Severity-driven complexity: Severity 1 (SyntaxError) ~2min avg. Severity 5 (SegFault) ~15min avg. Correlates with compile attempts, stack trace depth, and fix strategy complexity.
Context modifiers: On-call incidents resolve faster (pressure). Learning/tutorials take longest. Weekends slightly slower.
Language difficulty: Rust/C++ sessions take 1.4-1.6x longer. Python/JavaScript are fastest to debug.
Outcome realism: 88.6% fixed, 7.7% workaround, 2.8% escalated, 0.9% abandoned. Quality correlates with experience.
Event trajectories: Each session generates a realistic sequence of 3-200 events. Early = more errors/searches. Late = more tests/commits.
SUGGESTED DL TASKS
Time-to-fix regression: Predict resolution_time_seconds from first N events + session metadata.
Outcome classification: Predict fixed/escalated/abandoned from early trajectory patterns.
Next-event prediction: Given event history, predict next event_type.
Fix quality prediction: Predict fix_quality from debugging behavior.
Anomaly detection: Identify sessions that will regress or abandon before they do.
Sequence-to-sequence: Generate full event trajectory from session metadata (generative modeling).