scripts/train_all.sh

#!/usr/bin/env bash
set -euo pipefail

# ============================================================
# InsureOS Models — Train All Models Sequentially
# Designed for 16GB GPU — runs one model at a time
# ============================================================

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR="$(dirname "$SCRIPT_DIR")"
cd "$ROOT_DIR"

source .venv/bin/activate

echo "============================================"
echo " InsureOS Models — Full Training Pipeline"
echo " $(date)"
echo "============================================"

# Phase 0: Generate data
echo ""
echo "=== PHASE 0: Generating synthetic training data ==="
python -m data.generate_all

# Phase 1: QLoRA fine-tuning (InsureLLM-8B)
echo ""
echo "=== PHASE 1: QLoRA Fine-Tuning — Qwen3-8B ==="
python -m training.qlora_finetune \
    --base-model Qwen/Qwen3-8B \
    --dataset data/output/insurance_sft_10k.jsonl \
    --output-dir outputs/insurellm-8b \
    --epochs 3 \
    --batch-size 2 \
    --gradient-accumulation 8 \
    --learning-rate 2e-4 \
    --lora-rank 64 \
    --lora-alpha 128

# Phase 2: DPO reinforcement learning
echo ""
echo "=== PHASE 2: DPO Reinforcement Learning ==="
python -m training.dpo_train \
    --base-model Qwen/Qwen3-8B \
    --adapter-path outputs/insurellm-8b \
    --dataset data/output/insurance_dpo_5k.jsonl \
    --output-dir outputs/insurellm-8b-dpo \
    --epochs 1 \
    --batch-size 1 \
    --gradient-accumulation 16 \
    --learning-rate 5e-5 \
    --beta 0.1

# Phase 3: Distillation to smaller model
echo ""
echo "=== PHASE 3: Knowledge Distillation — Qwen3-4B ==="
python -m training.distill \
    --teacher-model Qwen/Qwen3-8B \
    --teacher-adapter outputs/insurellm-8b-dpo \
    --student-model Qwen/Qwen3-4B \
    --dataset data/output/insurance_sft_10k.jsonl \
    --output-dir outputs/insurellm-4b \
    --epochs 2 \
    --batch-size 2 \
    --gradient-accumulation 8

# Phase 4: Fraud detection model
echo ""
echo "=== PHASE 4: Fraud Detection Model ==="
python -m training.fraud_model \
    --dataset data/output/claims_tabular_50k.csv \
    --output-dir outputs/fraudnet

# Phase 5: Pricing GLM
echo ""
echo "=== PHASE 5: Pricing GLM ==="
python -m training.pricing_glm \
    --dataset data/output/claims_tabular_50k.csv \
    --output-dir outputs/pricing-glm

# Phase 6: Document classifier
echo ""
echo "=== PHASE 6: Document Classifier ==="
python -m training.doc_classifier \
    --dataset data/output/documents_10k.jsonl \
    --output-dir outputs/doc-classifier \
    --epochs 5 \
    --batch-size 16

# Phase 7: Insurance NER
echo ""
echo "=== PHASE 7: Insurance NER ==="
python -m training.ner_model \
    --dataset data/output/entities_8k.jsonl \
    --output-dir outputs/insure-ner \
    --epochs 5 \
    --batch-size 16

# Phase 8: Evaluation
echo ""
echo "=== PHASE 8: Running Evaluation Suite ==="
python -m evaluation.run_eval --all

echo ""
echo "============================================"
echo " Training Complete! $(date)"
echo " Models saved in outputs/"
echo "============================================"
ls -la outputs/