#!/bin/bash

# ==============================================================================
# Open-knowledge-Fine-tuning-transparency
# Training Script: Hermes Option D
# 
# EXECUTION INSTRUCTIONS:
# 1. Grant execution permissions:
#    chmod +x train_v4_D.sh
#
# 2. Run the script:
#    ./train_v4_D.sh
# ==============================================================================

# Configuration Parameters (Optimized for reproducibility)
# Ensure these match the intended dataset path in your local environment
DATASET_PATH="./datasets/dataset_30"
MODEL_PATH="./models/gemma-4-e2b-it-4bit"
OUTPUT_DIR="./weights/adapters30_v2"

echo "🚀 Starting training process..."
echo "⚙️ Config: Rank 8, Alpha 16, Iters 1000"

# 1. Prepare data
mkdir -p ./data
cp "$DATASET_PATH/train.jsonl" ./data/train.jsonl

# 2. Execute training
# Parameters are set to ensure convergence based on documented experiments
python3 -m mlx_vlm.lora \
      --model-path "$MODEL_PATH" \
      --dataset ./data \
      --batch-size 1 \
      --iters 1000 \
      --learning-rate 1e-5 \
      --gradient-accumulation-steps 16 \
      --steps-per-eval 50 \
      --val-batches 25 \
      --steps-per-save 50 \
      --output-path "$OUTPUT_DIR" \
      --grad-checkpoint \
      --lora-rank 8 \
      --lora-alpha 16 \
      --train-on-completions \
      --assistant-id 4368

echo "🎉 Training finished."
echo "✅ Adapter saved to: $OUTPUT_DIR"