This repo scaffolds data preparation + SFT + DPO training for a multi-turn Motivational Interviewing (MI) style coach model using Hugging Face Transformers + TRL.

• data/ — schema + a tiny synthetic example dataset (JSONL) to verify the pipeline.
• scripts/data_prep.py — normalize MI datasets (MI-TAGS / AnnoMI / MI-Dataset) into a common JSONL format.
• scripts/sft_train.py — Supervised fine-tuning with TRL SFTTrainer. Supports LoRA + 4-bit.
• scripts/dpo_train.py — Preference optimization with TRL DPOTrainer.
• scripts/infer_demo.py — Run inference with memory + MI persona prompt.
• scripts/metrics_mi.py — Simple MI-style behavioral metrics (coverage of open-question/reflect/affirm, etc.).
• configs/*.yaml — Example hyper-parameters.

⚠️ You must provide your own dataset paths for MI-TAGS / AnnoMI etc. The included example_mi_dialogs.jsonl is just for smoke tests (not for real training).

cd /mnt
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b -p /mnt/miniconda3
echo 'export PATH="/mnt/miniconda3/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc
conda create -n gptcoach python=3.10
conda activate gptcoach
pip install -r /mnt/gptcoaching_mi_training/requirements.txt

pip install -U transformers datasets accelerate trl peft bitsandbytes torch torchvision torchaudio
# If CUDA is not available, install CPU wheels for torch.

Each line is a dict:

{
  "dialog_id": "string",
  "turn_id": 7,
  "user_utt": "string",
  "coach_utt": "string",
  "mi_tags": ["open_question","reflection_simple","affirm"],
  "state_before": {},   # optional structured state (goal, barriers, wearable stats, etc.)
  "state_after":  {}    # optional updated state
}

You can include additional fields; unknown keys are ignored by the loader.

Set your dataset file paths and run:

python scripts/data_prep.py  --annomi_csv ./data/AnnoMI-full.csv --out_jsonl data/mi_unified_from_annomi_full.jsonl

python scripts/data_prep.py  --annomi_csv ./data/AnnoMI-simple.csv --out_jsonl data/mi_unified_from_annomi_simple.jsonl

python scripts/sft_train.py \
  --model_name_or_path Qwen/Qwen2.5-3B-Instruct \
  --train_file data/mi_unified_from_annomi_full.jsonl \
  --eval_file data/mi_unified_from_annomi_simple.jsonl \
  --output_dir outputs/qwen2p5-3b-mi-sft \
  --num_train_epochs 3 \
  --per_device_train_batch_size 20 \
  --lr 2e-5 \
  --eval_steps 200 \
  --save_steps 200 \
  --wandb --wandb_project mi-coach-sft --wandb_run_name qwen2p5_3b_sft \
  --bnb_4bit --lora

This part will generate both positive and negative samples (MI-style coaching response).

python scripts/make_dpo_prefs_v2.py \
  --sft_file data/mi_unified_from_annomi_full.jsonl \
  --out_file data/mi_prefs.jsonl \
  --seed 123 \
  --max_samples 5000

Prepare a JSONL with pairs of responses (chosen vs rejected) per context/turn.

python scripts/dpo_train.py \
  --model_name_or_path outputs/qwen2p5-3b-mi-sft/checkpoint-510 \
  --pref_file data/mi_prefs.jsonl \
  --output_dir runs/qwen2p5-3b-mi-dpo \
  --num_train_epochs 3 \
  --per_device_train_batch_size 1 \
  --lr 2e-5 \
  --logging_steps 10 \
  --save_steps 200 \
  --lora --bnb_4bit \
  --wandb --wandb_project mi-coach-dpo --wandb_run_name qwen2p5_3b_dpo

export HF_HOME=/mnt/.cache/huggingface
export TRANSFORMERS_CACHE=/mnt/.cache/huggingface/transformers

# merged model dir from your DPO (or SFT) run
export MODEL_PATH=/mnt/gptcoaching_mi_training/runs/qwen2p5-3b-mi-dpo-merged
uvicorn scripts.app_demo:app --host 0.0.0.0 --port 8000 --reload
# POST http://localhost:8000/chat
# body:
# {
#   "history": [{"user":"I want to be more active.","coach":"What matters most about being active for you?"}],
#   "user_msg":"I'm too busy this week."
# }

Kerrio.AI implements a Mayo Clinic-inspired 7-stage clinical journey for cognitive optimization. This is NOT a chatbot - it's a diagnostic-first digital cognitive clinic.

• Accurate diagnosis is the foundation of effective treatment
• Understanding is a prerequisite for permanent change
• Client History and Clinician's Notes are maintained separately

1. Registration - Client validated as invited guest
2. History Collection - Three Pillars (History, Psychology/Philosophy, Physiology)
3. Consultation - Clarify ambiguities, uncover blind spots
4. Diagnosis - Build Cognitive Wiring Map, explain WHY the problem exists
5. Proposal - Personalized treatment plan based on diagnosis
6. Treatment - Cognitive Rewiring Maps (Patent Pending)
7. Monitoring - Longitudinal progress assessment

# 1. Set environment variables
export HF_HOME=/mnt/.cache/huggingface
export TRANSFORMERS_CACHE=/mnt/.cache/huggingface/transformers
export MODEL_PATH=/mnt/gptcoaching_mi_training/runs/qwen2p5-3b-mi-dpo-merged

# 2. Start the server
uvicorn scripts.app_demo:app --host 0.0.0.0 --port 8000 --reload

# 3. Open browser
# http://localhost:8000/

• POST /api/chat - Send message and get AI response

  {"user_id": "demo_user", "user_msg": "I feel stuck in my career"}

• GET /api/journey/{user_id} - Get current journey status and stage
• POST /api/journey/advance - Advance to next stage (if requirements met)
• GET /api/journey/prompts/{user_id} - Get suggested prompts for current stage

• GET /api/journey/history/{user_id} - Get client's collected history across 3 pillars
• GET /api/journey/notes/{user_id} - Get clinician's notes (AI observations)

• GET /api/journey/diagnosis/{user_id} - Generate/retrieve diagnosis
• POST /api/journey/diagnosis/confirm/{user_id} - Confirm understanding of diagnosis
• GET /api/journey/treatment/{user_id} - Get treatment proposal with Cognitive Rewiring Map
• POST /api/journey/treatment/accept/{user_id} - Accept treatment plan
• POST /api/journey/treatment/progress/{user_id} - Update treatment progress

• GET /api/journey/videos - Get all educational video library
• GET /api/journey/videos/{video_id} - Get specific video details

• GET /api/journey/full-profile/{user_id} - Complete client profile with all data

• POST /api/cogmap - Build cognitive map from session
• GET /api/map/{user_id} - Get cognitive wiring map for user

The web UI (web/index.html) includes:

1. Journey Progress Bar - Shows current stage (Registration → Monitoring)
2. Chat Interface - Conversational interaction with Kerrio
3. Three Pillars Panel - View collected history across:
   • History Pillar (life events, patterns)
   • Psychology/Philosophy Pillar (beliefs, values)
   • Physiology Pillar (sleep, stress, health)
4. Diagnosis Panel - View:
   • Core Constraints
   • Bottlenecks
   • Root Causes
   • Explanation
   • Recommended Educational Videos
   • "I Understand My Diagnosis" confirmation button
5. Treatment Panel - View:
   • Current Wiring patterns
   • Target Wiring (desired state)
   • Rewiring Steps
   • Progress bar
   • "Accept Treatment Plan" button
6. Cognitive Map Visualization - Interactive graph with Cytoscape.js

# Run the kerrio_journey.py module directly for testing
python scripts/kerrio_journey.py

# This will:
# - Create a test profile
# - Show the stage-specific system prompt
# - Test the Diagnostic Engine
# - Display sample diagnosis output

┌─────────────────────────────────────────────────────────────────┐
│                        Web Interface                            │
│  (Chat, Journey Bar, Three Pillars, Diagnosis, Treatment)       │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                      FastAPI (app_demo.py)                       │
│  - /api/chat, /api/journey/*, /api/cogmap, /api/map/*           │
└─────────────────────────────────────────────────────────────────┘
                              │
          ┌───────────────────┼───────────────────┐
          ▼                   ▼                   ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ KerriJourney    │ │ DiagnosticEngine│ │ CognitiveRewiring│
│ Manager         │ │                 │ │ Engine           │
│ - 7 stages      │ │ - Root causes   │ │ - Rewiring maps  │
│ - 3 pillars     │ │ - Bottlenecks   │ │ - Treatment steps│
│ - Profile I/O   │ │ - Video recs    │ │                  │
└─────────────────┘ └─────────────────┘ └─────────────────┘
          │                   │                   │
          └───────────────────┴───────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                    LLM (Qwen/fine-tuned model)                   │
│            Stage-specific system prompts                         │
└─────────────────────────────────────────────────────────────────┘