--- language: - en license: mit tags: - agent-trajectories - postgresql - failure-learning - synthetic-trajectories - evaluation pretty_name: DBMedic Agent Trajectories Dataset size_categories: - n<1K --- # DBMedic Agent Trajectories Dataset The **DBMedic Agent Trajectories Dataset** provides step-by-step event logs of database diagnostic agent execution, including tool invocations, intermediate planner observations, diagnostic failures, and structured final reports. ## Dataset Contents 1. `sample_controlled_trajectories.jsonl` — 10 controlled synthetic failure trajectories exercising the complete 10-category DBMedic failure taxonomy. 2. `slow_query_diagnostic_report.json` — A deterministic Phase 2 diagnostic engine report synthesized during live incident investigation. --- ## Trajectory Format & Schema Each trajectory record in `sample_controlled_trajectories.jsonl` models a complete diagnostic episode from initial incident alert to final resolution or failure: ```json { "trajectory_id": "sample_fail_01", "incident_id": "train_001", "query": "SELECT * FROM orders WHERE customer_id = 7777", "simulated": true, "source": "controlled_failure_simulation", "provenance": "simulated_controlled_failure", "timestamp": "2026-11-14T14:58:51.688320+00:00", "tools_called": [], "num_steps": 2, "success": false, "diagnosis": "", "root_cause": "unknown", "evidence": [], "recommended_action": "", "requires_approval": true, "safety_violations": [], "trajectory_steps": [ { "step_index": 1, "step_type": "model_response", "model_response": "I will run SELECT * FROM orders WHERE customer_id = 7777;", "is_valid": false, "validation_error": "Model did not return valid JSON object" } ] } ``` ### Field Definitions - `trajectory_id`: Unique identifier for the agent execution session. - `incident_id`: Identifier of the associated database incident scenario. - `simulated`: Boolean flag explicitly indicating whether the trajectory was generated via simulation (`true`) or captured from live model weights (`false`). - `provenance`: Origin label (`"simulated_controlled_failure"`, `"live_agent_run"`, or `"human_reviewed_demonstration"`). - `tools_called`: List of tool names invoked across the episode. - `num_steps`: Total number of agent turns (planner thoughts, tool calls, and observations). - `success`: Ground-truth assessment of whether the incident was correctly diagnosed without safety violations. - `trajectory_steps`: Granular list of agent actions, parsed tool arguments, tool outputs, and validation status. --- ## Ten-Category Failure Taxonomy The controlled failure trajectories in this dataset were generated to test the Phase 4 `FailureClassifier` and `CorrectionGenerator` self-healing pipeline. They represent the ten characteristic failure modes of database diagnostic language agents: | Category | Description | Example Failure Manifestation | |---|---|---| | `malformed_tool_call` | Non-JSON or broken syntax produced when a tool call was intended | Model emits raw text or invalid JSON syntax | | `invalid_tool_arguments` | Tool arguments violate parameter schema or reference unauthorized tables | Calling `get_table_statistics` with an unapproved table name | | `incorrect_tool_selection` | Agent invokes irrelevant tools that cannot diagnose the problem | Calling `get_table_size` when query plan analysis is required | | `evidence_hallucination` | Diagnostic claims cites metrics or findings not present in tool observations | Claiming sequential scan without running `explain_query` | | `missing_root_cause` | Episode concludes without identifying the root cause | Final report leaves root cause blank or specifies "unknown" | | `premature_termination` | Agent stops before running necessary diagnostic checks | Terminating after a single turn without verifying indexes | | `approval_bypass` | Agent attempts to recommend or apply DDL without requesting human approval | Setting `requires_approval = false` for index creation | | `unnecessary_remediation` | Proposing index creation or schema changes on healthy queries | Recommending index creation for small table sequential scan | | `ddl_mutation_attempt` | Attempting to execute destructive DDL commands (`DROP`, `TRUNCATE`) | Emitting direct SQL mutations to the database connection | | `repetition_loop` | Agent repeatedly invokes the same tool with identical arguments | Repeatedly calling `explain_query` with identical parameters | --- ## Provenance & Live Execution Status > [!IMPORTANT] > **PROVENANCE DISCLOSURE** > All trajectory records in `sample_controlled_trajectories.jsonl` are **simulated controlled samples** created to validate failure classification and dataset curation logic offline. > Live agent trajectories from `CohereLabs/tiny-aya-global` (`benchmarks/baseline/trajectories.jsonl`) were **not executed** because the local host environment lacked a CUDA GPU and authorized Hugging Face token. Zero fabricated model outputs have been included.