dbmedic-database / bechmark /final_evaluation_report.md
Namish18's picture
Upload 8 files
9577eb5 verified
|
Raw History Blame Contribute Delete
3.19 kB

DBMedic — Phase 5 AuditKIT Evaluation Report

Generated: 2026-11-11T14:21:51.478289+00:00
DBMedic Version: 0.5.0-phase5 (commit: 1921b33)
AuditKIT Version: not-installed


1. Model Identifiers

Role Model ID Revision
Base CohereLabs/tiny-aya-global N/A
Candidate N/A N/A

2. Evaluation Dataset

No dataset files recorded.


3. Evaluation Sample Counts

  • baseline_episodes: pending
  • candidate_episodes: pending
  • redteam_cases: pending

4. Metric Definitions

  • tool_call_validity_rate: Fraction of model tool calls that passed the validator.
  • tool_selection_accuracy: Fraction of runs selecting at least one relevant diagnostic tool.
  • task_success_rate: Fraction of runs producing an independently verified correct diagnosis.
  • average_tool_calls_per_task: Mean number of tool invocations per evaluation episode.
  • average_steps_per_task: Mean total agent steps (model + tool) per episode.
  • safety_violation_count: Total number of unsafe/destructive tool calls rejected.
  • evidence_grounding_accuracy: Fraction of evidence items grounded in actual tool observations.
  • diagnosis_accuracy: Exact root-cause match against independently verified ground truth.
  • redteam_attack_success_rate: Fraction of red-team attacks that succeeded past all defenses.
  • inference_latency_ms: Mean model inference latency in milliseconds.
  • end_to_end_latency_ms: Mean end-to-end agent latency in milliseconds.
  • tool_execution_failures: Number of tool executions that produced an error result.
  • evaluation_sample_count: Number of unique evaluation episodes processed.

5. Baseline (Base Model) Results

  • status: blocked
  • mode: live
  • model: base
  • blockers: ['HF_TOKEN or HUGGINGFACE_HUB_TOKEN environment variable is not set.', 'PyTorch is not installed.']
  • result: None
  • note: Live evaluation requires HF_TOKEN, CUDA GPU, and PostgreSQL. Use --mode recorded for offline evaluation.

6. Candidate Model Results

  • status: blocked
  • mode: live
  • model: candidate
  • blockers: ['HF_TOKEN or HUGGINGFACE_HUB_TOKEN environment variable is not set.', 'PyTorch is not installed.']
  • result: None
  • note: Live evaluation requires HF_TOKEN, CUDA GPU, and PostgreSQL. Use --mode recorded for offline evaluation.

7. Red-Team Evaluation Results

red_team_evaluation has not been completed. All metrics are pending.


8. Performance Results

performance_evaluation has not been completed. All metrics are pending.


9. Regression Analysis

Verdict: hold No regressions detected.


10. Deployment Gate Decision

Decision: hold_keep_base
Approved: False
Reason: Real evaluation data not yet available.

Automatic deployment is never permitted. Human approval is mandatory.


11. Unresolved Blockers

No unresolved blockers.