Problem

In a Databricks AI pipeline, mlflow.evaluate and LLM-as-Judge outputs may attach free-form issue labels such as faithfulness, groundedness, or prompt injection to each response. Write idiomatic Python to normalize these labels into a clean, deterministic list suitable for downstream use in Mosaic AI and Model Serving logs.

Implement a function that takes a list of raw label strings and returns a sorted list of unique normalized labels.

Formal Specification

• Input: labels, a list of strings.
• Output: a list of strings where each label is:
  1. lowercased,
  2. trimmed of leading/trailing whitespace,
  3. internal runs of spaces, underscores, or hyphens collapsed into a single hyphen,
  4. stripped of all other non-alphanumeric characters except hyphens,
  5. deduplicated,
  6. returned in ascending lexicographic order.
• Ignore labels that become empty after normalization.

Examples

Example 1 Input: labels = [" Faithfulness ", "groundedness", "Prompt Injection", "prompt_injection"] Output: ["faithfulness", "groundedness", "prompt-injection"]

Explanation: Prompt Injection and prompt_injection normalize to the same canonical label.

Example 2 Input: labels = ["DBRX", "dbrx!!", "", " ", "model-serving"] Output: ["dbrx", "model-serving"]

Explanation: Empty values are ignored, punctuation is removed, and duplicates are collapsed.

Constraints

• 1 <= len(labels) <= 10^4
• 0 <= len(labels[i]) <= 200
• Labels may contain letters, digits, spaces, underscores, hyphens, and punctuation.