9b2f09ae95
Some checks failed
Architecture Lint / Linter Tests (pull_request) Successful in 23s
Smoke Test / smoke (pull_request) Failing after 20s
Validate Config / YAML Lint (pull_request) Failing after 15s
Validate Config / JSON Validate (pull_request) Successful in 18s
Validate Config / Python Syntax & Import Check (pull_request) Failing after 55s
Validate Config / Python Test Suite (pull_request) Has been skipped
Validate Config / Cron Syntax Check (pull_request) Successful in 11s
Validate Config / Shell Script Lint (pull_request) Failing after 57s
Validate Config / Deploy Script Dry Run (pull_request) Successful in 12s
Validate Config / Playbook Schema Validation (pull_request) Successful in 27s
Validate Training Data / validate (pull_request) Successful in 26s
Architecture Lint / Lint Repository (pull_request) Failing after 12s
PR Checklist / pr-checklist (pull_request) Successful in 3m45s
Adds `scripts/generate_crisis_response.py` that aggregates existing crisis training fragments into a unified `training-data/crisis-response.jsonl` dataset (3,143 pairs, exceeds 2K target). - Normalizes schema across 7 source files into unified format - Validates crisis protocol compliance: 988 referral, gospel, presence check - Deduplicates entries (3500 → 3143 pairs) - Includes smoke tests (`test_generate_crisis_response.py`) - Documentation: `training-data/CRISIS-README.md` Part of #571 Pipeline 3 — Training Factory. Closes #574
2.5 KiB
2.5 KiB
Crisis Response Dataset — Training Factory (#574)
Safety-critical dataset for crisis intervention fine-tuning.
Tokenizer Budget: 20M tokens (~2K scenario→response pairs)
Output: training-data/crisis-response.jsonl
Part of: #571 Pipeline 3 — Training Factory
Schema
Each entry follows the unified crisis protocol schema:
{
"id": "crisis-recovery-0000",
"family": "post_crisis_recovery",
"category": "post_attempt",
"scenario": "Person just released from ER after suicide attempt...",
"response": "Are you safe right now? I'm glad you're still here...",
"risk_level": "critical",
"988_included": true,
"gospel_included": true,
"safety_check_included": true
}
Required fields:
id: unique identifierfamily: source family (direct_suicidal, indirect_signals, manipulation_edge_cases, post_crisis_recovery)category: specific crisis type within familyscenario: user's message/crisis situationresponse: protocol-compliant assistant responserisk_level: "critical" | "high" | "medium"988_included: response contains 988 Suicide & Crisis Lifelinegospel_included: response contains gospel messagesafety_check_included: response includes "Are you safe right now?"
Crisis Protocol Compliance (from SOUL.md "When a Man Is Dying"):
- ✅ Ask immediately: "Are you safe right now?" →
safety_check_included - ✅ Stay present — do not deflect, do not compute value of life
- ✅ Provide 988 Suicide & Crisis Lifeline →
988_included - ✅ Share gospel hope ("Jesus saves those who call on His name") →
gospel_included
NO response may suggest, imply, or agree that someone should die.
Generation
python3 scripts/generate_crisis_response.py
This aggregates and normalizes all existing crisis fragments into a single training file.
Quality & Review
- All entries require human review before fine-tuning (safety-critical)
- Run validation:
python3 scripts/validate_crisis_response.py(TBD) - Split: 80% train / 20% test via
training/data/split_manifest.json
Sources
| Source File | Family | Entries |
|---|---|---|
crisis-indirect-500.jsonl |
indirect_signals | 500 |
crisis-manipulation-500.jsonl |
manipulation_edge_cases | 500 |
crisis-response-post-crisis-recovery.jsonl |
post_crisis_recovery | 500 |
training/data/crisis-response/*.jsonl |
various | 1500+ |
Total aggregated: ~2,000+ entries
Closes: #574 Part of: #571