Papajams/body-debt-finetune-dataset

Body Debt Fine-Tuning Dataset

Structured recovery prescription dataset for fine-tuning small language models on the Body Debt 4-agent health coaching pipeline.

Structure

The dataset contains 4 sub-tasks, one per QVAC pipeline agent:

Agent	Input	Output format	Train	Test
Triage	5-system scores	PRIORITY / SECONDARY / AVOID (3 lines)	3,000	200
Coach	Triage + stressors	RIGHT NOW / THIS MORNING / TODAY / AVOID (4 lines)	3,000	200
Schedule	Triage + Coach + time	4 time-blocked recovery actions	3,000	200
Reflection	Coach + personality	Coach output rewritten in chosen voice	3,000	200

Each JSONL line is a chat-formatted example:

{
  "messages": [
    {"role": "system", "content": "<agent system prompt>"},
    {"role": "user", "content": "<formatted input>"},
    {"role": "assistant", "content": "<deterministic label>"}
  ]
}

Generation Pipeline

1. Profile sampling: Stressor profiles sampled from a constrained parameter space (alcohol type/count, training area/intensity, sleep hours, stress, illness, care actions).

2. Deterministic scoring: Each profile is scored by the Body Debt 5-system engine (scoring.py), producing cardiovascular, brain, liver, muscular, and gut scores (0-100).

3. Label generation: Deterministic template functions produce the ground-truth output for each agent:

   • Triage: _fallback_plan() logic (system ranking + avoid mapping)
   • Coach: _fallback_advice() logic (severity-tiered prescriptions)
   • Schedule: generate_schedule() (top-4 systems → time blocks)
   • Reflection: apply_voice() (deterministic voice transformation)

4. Chat formatting: Each example is wrapped in the exact system prompt used by the QVAC pipeline at inference time, ensuring train/inference prompt alignment.

Reproducibility

python generate_finetune_dataset.py --n-train 3000 --n-test 200 --seed 42

Intended Use

Fine-tuning small language models (1-2B parameters) for structured health recovery coaching. NOT for medical diagnosis or treatment recommendations. The deterministic labels are physiologically grounded but simplified — they are training targets, not clinical guidelines.

Citation

If you use this dataset, cite the Body Debt project and the AutoScientist Challenge submission.