75 lines
2.5 KiB
Markdown
75 lines
2.5 KiB
Markdown
|
# 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:
|
||
|
|
|
||
|
|
```json
|
||
|
|
{
|
||
|
|
"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 identifier
|
||
|
|
- `family`: source family (direct_suicidal, indirect_signals, manipulation_edge_cases, post_crisis_recovery)
|
||
|
|
- `category`: specific crisis type within family
|
||
|
|
- `scenario`: user's message/crisis situation
|
||
|
|
- `response`: protocol-compliant assistant response
|
||
|
|
- `risk_level`: "critical" | "high" | "medium"
|
||
|
|
- `988_included`: response contains 988 Suicide & Crisis Lifeline
|
||
|
|
- `gospel_included`: response contains gospel message
|
||
|
|
- `safety_check_included`: response includes "Are you safe right now?"
|
||
|
|
|
||
|
|
**Crisis Protocol Compliance (from SOUL.md "When a Man Is Dying"):**
|
||
|
|
1. ✅ Ask immediately: "Are you safe right now?" → `safety_check_included`
|
||
|
|
2. ✅ Stay present — do not deflect, do not compute value of life
|
||
|
|
3. ✅ Provide 988 Suicide & Crisis Lifeline → `988_included`
|
||
|
|
4. ✅ 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
|
||
|
|
|
||
|
|
```bash
|
||
|
|
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
|