Create safety_config.json
Browse files- safety_config.json +124 -0
safety_config.json
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"version": "1.0",
|
| 3 |
+
"safety_settings": {
|
| 4 |
+
"enabled": true,
|
| 5 |
+
"strictness_level": "moderate",
|
| 6 |
+
"categories": {
|
| 7 |
+
"hate_speech": {
|
| 8 |
+
"enabled": true,
|
| 9 |
+
"threshold": 0.7,
|
| 10 |
+
"action": "filter"
|
| 11 |
+
},
|
| 12 |
+
"violence": {
|
| 13 |
+
"enabled": true,
|
| 14 |
+
"threshold": 0.75,
|
| 15 |
+
"action": "filter"
|
| 16 |
+
},
|
| 17 |
+
"self_harm": {
|
| 18 |
+
"enabled": true,
|
| 19 |
+
"threshold": 0.65,
|
| 20 |
+
"action": "block"
|
| 21 |
+
},
|
| 22 |
+
"sexual_content": {
|
| 23 |
+
"enabled": true,
|
| 24 |
+
"threshold": 0.7,
|
| 25 |
+
"action": "filter"
|
| 26 |
+
},
|
| 27 |
+
"illegal_activities": {
|
| 28 |
+
"enabled": true,
|
| 29 |
+
"threshold": 0.8,
|
| 30 |
+
"action": "block"
|
| 31 |
+
},
|
| 32 |
+
"child_safety": {
|
| 33 |
+
"enabled": true,
|
| 34 |
+
"threshold": 0.5,
|
| 35 |
+
"action": "block"
|
| 36 |
+
},
|
| 37 |
+
"privacy_violation": {
|
| 38 |
+
"enabled": true,
|
| 39 |
+
"threshold": 0.7,
|
| 40 |
+
"action": "warn"
|
| 41 |
+
},
|
| 42 |
+
"misinformation": {
|
| 43 |
+
"enabled": true,
|
| 44 |
+
"threshold": 0.75,
|
| 45 |
+
"action": "warn"
|
| 46 |
+
}
|
| 47 |
+
},
|
| 48 |
+
"content_filtering": {
|
| 49 |
+
"profanity_filter": {
|
| 50 |
+
"enabled": true,
|
| 51 |
+
"severity": "moderate"
|
| 52 |
+
},
|
| 53 |
+
"pii_detection": {
|
| 54 |
+
"enabled": true,
|
| 55 |
+
"types": [
|
| 56 |
+
"email",
|
| 57 |
+
"phone",
|
| 58 |
+
"ssn",
|
| 59 |
+
"credit_card",
|
| 60 |
+
"address"
|
| 61 |
+
]
|
| 62 |
+
},
|
| 63 |
+
"toxicity_detection": {
|
| 64 |
+
"enabled": true,
|
| 65 |
+
"threshold": 0.7
|
| 66 |
+
}
|
| 67 |
+
},
|
| 68 |
+
"rate_limiting": {
|
| 69 |
+
"enabled": true,
|
| 70 |
+
"requests_per_minute": 60,
|
| 71 |
+
"tokens_per_minute": 90000,
|
| 72 |
+
"concurrent_requests": 10
|
| 73 |
+
},
|
| 74 |
+
"output_validation": {
|
| 75 |
+
"max_tokens": 4096,
|
| 76 |
+
"min_tokens": 1,
|
| 77 |
+
"blocked_patterns": [
|
| 78 |
+
"(?i)(password|api[_-]?key|secret|token)\\s*[:=]\\s*['\"]?[a-zA-Z0-9]{8,}",
|
| 79 |
+
"(?i)-----BEGIN.*PRIVATE KEY-----"
|
| 80 |
+
],
|
| 81 |
+
"requires_citation": false
|
| 82 |
+
}
|
| 83 |
+
},
|
| 84 |
+
"guardrails": {
|
| 85 |
+
"input_validation": {
|
| 86 |
+
"max_prompt_length": 200000,
|
| 87 |
+
"max_images_per_request": 10,
|
| 88 |
+
"allowed_image_formats": ["jpg", "jpeg", "png", "webp", "gif"],
|
| 89 |
+
"max_image_size_mb": 20
|
| 90 |
+
},
|
| 91 |
+
"behavior_policies": {
|
| 92 |
+
"no_impersonation": true,
|
| 93 |
+
"no_financial_advice": false,
|
| 94 |
+
"no_medical_diagnosis": false,
|
| 95 |
+
"no_legal_advice": false,
|
| 96 |
+
"require_factual_grounding": true,
|
| 97 |
+
"cite_sources_when_uncertain": true
|
| 98 |
+
},
|
| 99 |
+
"ethical_constraints": {
|
| 100 |
+
"refuse_harmful_requests": true,
|
| 101 |
+
"refuse_bias_amplification": true,
|
| 102 |
+
"refuse_manipulation": true,
|
| 103 |
+
"refuse_deception": true,
|
| 104 |
+
"transparency_about_capabilities": true,
|
| 105 |
+
"acknowledge_limitations": true
|
| 106 |
+
}
|
| 107 |
+
},
|
| 108 |
+
"monitoring": {
|
| 109 |
+
"log_safety_events": true,
|
| 110 |
+
"log_filtered_content": true,
|
| 111 |
+
"alert_on_repeated_violations": true,
|
| 112 |
+
"track_user_patterns": false,
|
| 113 |
+
"anonymize_logs": true
|
| 114 |
+
},
|
| 115 |
+
"emergency_shutdown": {
|
| 116 |
+
"enabled": true,
|
| 117 |
+
"triggers": [
|
| 118 |
+
"repeated_harmful_content",
|
| 119 |
+
"attempted_jailbreak",
|
| 120 |
+
"system_abuse"
|
| 121 |
+
],
|
| 122 |
+
"cooldown_period_minutes": 15
|
| 123 |
+
}
|
| 124 |
+
}
|