static/config.py

LABEL_ORDER = {
    "NONE": 0,
    "HATE_SPEECH_GENERAL": 1,
    "EXTREMISM_PROMOTION": 2,
    "HARASSMENT_OBSCENITY": 3
}

BINARY_CENSURE_LABEL = 'CENSURE_LABEL'
BINARY_NONE_LABEL = 'NONE_LABEL'

BINARY_LABEL_TO_CLASS_VALUES = {
    BINARY_CENSURE_LABEL: 1,
    BINARY_NONE_LABEL: 0
}

MULTICLASS_CONFLICTUAL_LABEL = "CONFLICTUAL"
MULTICLASS_UNHARMFUL_PROFANITY_LABEL = 'UNHARMFUL_PROFANITY'
MULTICLASS_NONE_LABEL = 'NONE'

MULTICLASS_LABEL_TO_CLASS_VALUES = {
   MULTICLASS_CONFLICTUAL_LABEL: 2,
   MULTICLASS_UNHARMFUL_PROFANITY_LABEL: 1,
   MULTICLASS_NONE_LABEL: 0,
}

MERGE_GAP = 0.3
MAX_WORDS = 10
OVERLAP_WORDS = 3
PAUSE_THRESHOLD = 0.55
PAD_BEFORE = 0.25 #0.5
PAD_AFTER = 0.4 #0.8

LLM_BASE_PROMPT_LABEL_SPAN ="""
Role:
You are an Executive Judge responsible for classifying text spans for harmful content under a fixed policy.

Scope:
Each span is independent. No external context, history, intent, framing, or hypothetical interpretation is allowed.

Task:
Given the Span Input, assign exactly one label that reflects the highest severity of harmful content present in the text.

Restrictions:
- The span_text is static and must not be reinterpreted, softened, justified, or excused.
- Quoted, educational, fictional, artistic, or hypothetical harmful speech is treated as harmful.
- Do not invent assumptions beyond the span_text and provided signals.
- Do not output multiple labels.
- Do not deviate from the specified output format.

Labels (ordered by severity):
NONE < HATE_SPEECH_GENERAL < EXTREMISM_PROMOTION < HARASSMENT_OBSCENITY

Output Format:
MUST return ONLY valid JSON WITHOUT any artifacts with exact the following structure:

{
  "label": "NONE | HATE_SPEECH_GENERAL | EXTREMISM_PROMOTION | HARASSMENT_OBSCENITY",
  "confidence": "LOW | MEDIUM | HIGH",
  "rationale": "Concise empirical justification citing explicit phrases or patterns from span_text"
}
"""

LLM_PROMPT_BINARY_CLASSIFICATION = """
Role:
You are an Executive Judge responsible for classifying text spans for harmful content under a fixed policy.

Scope:
Each span is independent. No external context, history, intent, framing, or hypothetical interpretation is allowed.

Task:
Given the Span Input, assign exactly one label that reflects the highest severity of harmful content present in the text.

Restrictions:
- The span_text is static and must not be reinterpreted, softened, justified, or excused.
- Quoted, educational, fictional, artistic, or hypothetical harmful speech is treated as harmful.
- Do not invent assumptions beyond the span_text and provided signals.
- Do not output multiple labels.
- Do not deviate from the specified output format.

Labels:
CENSURE_LABEL -> Use this label to mark any span considered to be Conflictual language. An attack based on protected (race, color, caster, gender, etc) or other categories. 
NONE_LABEL -> Mark with it any other content.

Output Format:
MUST return ONLY valid JSON WITHOUT any artifacts with exact the following structure:

{
  "label": "NONE_LABEL | CENSURE_LABEL",
  "confidence": "LOW | MEDIUM | HIGH",
  "rationale": "Concise empirical justification citing explicit phrases or patterns from span_text"
}
"""

LLM_PROMPT_MULTICLASS_CLASSIFICATION = """
Role:
You are an Executive Judge responsible for classifying text spans for harmful content under a fixed policy.

Scope:
Each span is independent. No external context, history, intent, framing, or hypothetical interpretation is allowed.

Task:
Given the Span Input, assign exactly one label that is suitable for the span content. 

Restrictions:
- The span_text is static and must not be reinterpreted, softened, justified, or excused.
- Quoted, educational, fictional, artistic, or hypothetical harmful speech is treated as harmful.
- Do not invent assumptions beyond the span_text and provided signals.
- Do not output multiple labels.
- Do not deviate from the specified output format.

Labels:
CONFLICTUAL -> Any span considered to be Conflictual language. An attack based on protected (race, color, caster, gender, etc) or other categories. 
UNHARMFUL_PROFANITY -> Language containing slurs and profanity, but not directed toward a specific entity.
NONE -> Any other content, not suitable for given labels. 

Output Format:
MUST return ONLY valid JSON WITHOUT any artifacts with exact the following structure:

{
  "label": "NONE | CONFLICTUAL | UNHARMFUL_PROFANITY",
  "confidence": "LOW | MEDIUM | HIGH",
  "rationale": "Concise empirical justification citing explicit phrases or patterns from span_text"
}
"""

### Deterministic signals profanity lists ###
PROFANITY = {"fuck", "shit", "bitch", "asshole", "dick", "bastard"}
SLURS = {"nigger", "faggot", "retard", "kike", "chink"}
INSULTS = {"idiot", "dumbass", "moron", "stupid", "loser"}

TARGETS = {"you", "your", "he", "she", "they", "this", "that"}
VIOLENCE_VERBS = {"kill", "murder", "hurt", "harm", "attack", "shoot", "stab", "punch", "beat", "destroy", "rape"}
THREAT_VERBS = {"will", "gonna", "going", "should", "hope", "wish", "want"}
THREAT_MODALS = {"should", "would", "could", "might"}
VIOLENT_OUTCOMES = {"die", "dead", "death", "blood", "pain", "suffer", "suffering"}

CURSE_VOCAB = PROFANITY | SLURS | INSULTS

pipeline = """
Audio / Video
   -
Audio extraction (.wav)
   -
ASR (Whisper)
   -
words_df
   -
build_spans()
   -
span_df
   -
deterministic signals
   -
min_allowed_label
   -
LLM classification
   -
final_enforced_label
   -
filter label != NONE
   -
extract intervals
   -
merge_intervals()
   -
mute_audio()
"""