{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "3mmJUlxcRpiU"
      },
      "outputs": [],
      "source": [
        "# Colab cell 1: Install dependencies\n",
        "!pip install --quiet \\\n",
        "  transformers accelerate peft datasets \\\n",
        "  bitsandbytes huggingface_hub \\\n",
        "  pymupdf"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab cell 2: Log in to Hugging Face\n",
        "from huggingface_hub import notebook_login\n",
        "notebook_login()\n",
        "# This will prompt you to paste a Hugging Face access token."
      ],
      "metadata": {
        "id": "H3KO-a_dSod6"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab cell 3: Mount your Drive (if PDFs are there)\n",
        "from google.colab import drive\n",
        "drive.mount('/content/drive')"
      ],
      "metadata": {
        "id": "oMaXhRSCVu51"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab cell 4: Extract text from all PDFs\n",
        "import fitz  # PyMuPDF\n",
        "import os\n",
        "\n",
        "PDF_DIR = \"/content/drive/MyDrive/GodeusAI-DatasetPDF\"\n",
        "OUTPUT_TXT = \"/content/all_text.txt\"\n",
        "\n",
        "with open(OUTPUT_TXT, \"w\", encoding=\"utf-8\") as fout:\n",
        "    for fname in os.listdir(PDF_DIR):\n",
        "        if fname.lower().endswith(\".pdf\"):\n",
        "            doc = fitz.open(os.path.join(PDF_DIR, fname))\n",
        "            for page in doc:\n",
        "                fout.write(page.get_text())\n",
        "            doc.close()\n",
        "print(\"✅ Extracted text from PDFs to\", OUTPUT_TXT)"
      ],
      "metadata": {
        "id": "j_pVh2BcWn2b"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab cell 5: Chunk & format into JSONL\n",
        "import tiktoken  # or use your tokenizer for approximate token counts\n",
        "import json\n",
        "\n",
        "def chunk_text(text, max_tokens=512, overlap=50):\n",
        "    # simple whitespace split + sliding window\n",
        "    words = text.split()\n",
        "    chunks = []\n",
        "    i = 0\n",
        "    while i < len(words):\n",
        "        chunk = words[i : i + max_tokens]\n",
        "        chunks.append(\" \".join(chunk))\n",
        "        i += max_tokens - overlap\n",
        "    return chunks\n",
        "\n",
        "# Read and chunk\n",
        "with open(OUTPUT_TXT, \"r\", encoding=\"utf-8\") as fin:\n",
        "    text = fin.read()\n",
        "\n",
        "chunks = chunk_text(text, max_tokens=512, overlap=50)\n",
        "\n",
        "# Build instruction–response pairs (empty input, you can customize)\n",
        "records = []\n",
        "for chunk in chunks:\n",
        "    records.append({\n",
        "        \"instruction\": \"Based on this teaching, explain the key insight in a concise coach‑style voice.\",\n",
        "        \"input\": chunk,\n",
        "        \"output\": \"\"  # leave blank for self‑supervised teaching; or fill with human summaries\n",
        "    })\n",
        "\n",
        "# Save JSONL\n",
        "import pathlib\n",
        "out_path = pathlib.Path(\"/content/godeusai_instruct.jsonl\")\n",
        "with out_path.open(\"w\", encoding=\"utf-8\") as fout:\n",
        "    for rec in records:\n",
        "        fout.write(json.dumps(rec) + \"\\n\")\n",
        "print(\"✅ Wrote\", len(records), \"records to\", out_path)"
      ],
      "metadata": {
        "id": "w_oi1OEnXgv2"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from huggingface_hub import login\n",
        "login()  # paste your token when prompted\n"
      ],
      "metadata": {
        "id": "IbsSoQxHZ789"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab cell 6: Load model in 4‑bit + LoRA configuration\n",
        "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
        "from peft import LoraConfig, get_peft_model, TaskType\n",
        "\n",
        "MODEL_NAME = \"mistralai/Mistral-7B-v0.1\"\n",
        "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True)\n",
        "\n",
        "# 4‑bit quantization config\n",
        "bnb_config = BitsAndBytesConfig(\n",
        "    load_in_4bit=True,\n",
        "    bnb_4bit_quant_type=\"nf4\",\n",
        "    bnb_4bit_use_double_quant=True,\n",
        "    bnb_4bit_compute_dtype=\"bfloat16\"\n",
        ")\n",
        "\n",
        "model = AutoModelForCausalLM.from_pretrained(\n",
        "    MODEL_NAME,\n",
        "    device_map=\"auto\",\n",
        "    quantization_config=bnb_config\n",
        ")\n",
        "\n",
        "# LoRA adapter setup\n",
        "peft_config = LoraConfig(\n",
        "    task_type=TaskType.CAUSAL_LM,\n",
        "    inference_mode=False,\n",
        "    r=32,             # adapter rank—controls capacity to learn style\n",
        "    lora_alpha=16,\n",
        "    lora_dropout=0.05\n",
        ")\n",
        "model = get_peft_model(model, peft_config)\n",
        "def count_trainable_params(model):\n",
        "    return sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
        "\n",
        "print(\"✅ Model + LoRA ready. Trainable params:\", count_trainable_params(model))\n"
      ],
      "metadata": {
        "id": "2z9pACO5Y6OP"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab Cell 7: Safely load local JSONL + tokenize\n",
        "import json\n",
        "from datasets import Dataset\n",
        "\n",
        "# Assign pad_token (Mistral doesn't define one by default)\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "# Load JSONL into memory\n",
        "with open(\"/content/godeusai_instruct.jsonl\", \"r\") as f:\n",
        "    raw_data = [json.loads(line) for line in f]\n",
        "\n",
        "# Convert to Hugging Face Dataset\n",
        "ds = Dataset.from_list(raw_data)\n",
        "ds = ds.train_test_split(test_size=0.05)\n",
        "\n",
        "# Tokenization logic\n",
        "def tokenize_fn(example):\n",
        "    prompt = (\n",
        "        f\"### Instruction:\\n{example['instruction']}\\n\"\n",
        "        f\"### Input:\\n{example['input']}\\n\"\n",
        "        f\"### Response:\\n{example['output']}\"\n",
        "    )\n",
        "    tokens = tokenizer(\n",
        "        prompt,\n",
        "        truncation=True,\n",
        "        max_length=600,\n",
        "        padding=\"max_length\"\n",
        "    )\n",
        "    return tokens  # ❌ Do not add \"labels\"\n",
        "\n",
        "\n",
        "# Apply tokenization\n",
        "# **NOTE** batched=False\n",
        "tokenized = ds.map(\n",
        "    tokenize_fn,\n",
        "    batched=False,\n",
        "    remove_columns=ds[\"train\"].column_names\n",
        ")\n",
        "print(f\"✅ Tokenization done. Example input_ids length: {len(tokenized['train'][0]['input_ids'])}\")\n"
      ],
      "metadata": {
        "id": "MPDZW2IedGJM"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Colab Cell 8: Train with robust filtering + custom collator\n",
        "\n",
        "import torch\n",
        "from torch.nn.utils.rnn import pad_sequence\n",
        "from transformers import Trainer, TrainingArguments\n",
        "\n",
        "# 1) Filter out empty examples\n",
        "def is_valid(ex): return isinstance(ex[\"input_ids\"], list) and len(ex[\"input_ids\"]) > 0\n",
        "\n",
        "for split in [\"train\", \"test\"]:\n",
        "    before = len(tokenized[split])\n",
        "    tokenized[split] = tokenized[split].filter(is_valid)\n",
        "    after = len(tokenized[split])\n",
        "    print(f\"✅ {split}: {before} → {after}\")\n",
        "\n",
        "# 2) Custom collator (CPU tensors only)\n",
        "def causal_collator(batch):\n",
        "    input_ids      = [torch.tensor(ex[\"input_ids\"],      dtype=torch.long) for ex in batch]\n",
        "    attention_mask = [torch.tensor(ex[\"attention_mask\"], dtype=torch.long) for ex in batch]\n",
        "    input_ids      = pad_sequence(input_ids,      batch_first=True, padding_value=tokenizer.eos_token_id)\n",
        "    attention_mask = pad_sequence(attention_mask, batch_first=True, padding_value=0)\n",
        "    labels = input_ids.clone()\n",
        "    return {\n",
        "        \"input_ids\":      input_ids,\n",
        "        \"attention_mask\": attention_mask,\n",
        "        \"labels\":         labels,\n",
        "    }\n",
        "\n",
        "# 3) TrainingArguments with pin_memory disabled\n",
        "training_args = TrainingArguments(\n",
        "    output_dir=\"/content/GodeusAI_lora\",\n",
        "    per_device_train_batch_size=1,\n",
        "    gradient_accumulation_steps=8,\n",
        "    num_train_epochs=3,\n",
        "    logging_steps=50,\n",
        "    save_strategy=\"epoch\",\n",
        "    learning_rate=2e-4,\n",
        "    fp16=True,\n",
        "    optim=\"paged_adamw_32bit\",\n",
        "    push_to_hub=True,\n",
        "    report_to=\"none\",\n",
        "    dataloader_pin_memory=False,  # disable pinning\n",
        ")\n"
      ],
      "metadata": {
        "id": "WmAvAV4reqGv"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Save the adapter\n",
        "model.save_pretrained(\"/content/GodeusAI_adapter\")\n"
      ],
      "metadata": {
        "id": "xaF33MpOraS4"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from huggingface_hub import upload_folder\n",
        "\n",
        "upload_folder(\n",
        "    folder_path=\"/content/GodeusAI_adapter\",\n",
        "    repo_id=\"yadnik/GodeusAI-v1\",\n",
        "    repo_type=\"model\"\n",
        ")\n"
      ],
      "metadata": {
        "id": "W3J0RktItdYy"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
        "from peft import PeftModel\n",
        "\n",
        "MODEL_NAME   = \"mistralai/Mistral-7B-v0.1\"\n",
        "ADAPTER_REPO = \"yadnik/GodeusAI-v1\"\n",
        "\n",
        "# 1) 4‑bit quantization config (from transformers)\n",
        "bnb_config = BitsAndBytesConfig(\n",
        "    load_in_4bit=True,\n",
        "    bnb_4bit_quant_type=\"nf4\",\n",
        "    bnb_4bit_use_double_quant=True,\n",
        "    bnb_4bit_compute_dtype=\"bfloat16\"\n",
        ")\n",
        "\n",
        "# 2) Tokenizer\n",
        "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True)\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "# 3) Load base model in 4‑bit\n",
        "base_model = AutoModelForCausalLM.from_pretrained(\n",
        "    MODEL_NAME,\n",
        "    device_map=\"auto\",\n",
        "    quantization_config=bnb_config\n",
        ")\n",
        "\n",
        "# 4) Attach your LoRA adapter\n",
        "model = PeftModel.from_pretrained(\n",
        "    base_model,\n",
        "    ADAPTER_REPO,\n",
        "    device_map=\"auto\"\n",
        ")\n",
        "\n",
        "# 5) Inference helper\n",
        "def ask_discepline(prompt: str, max_new_tokens: int = 200):\n",
        "    persona = (\n",
        "        \"You are Godeus AI—a compassionate, omniscient guide inspired by the timeless wisdom of spiritual figures and universal truths. You embody the serene, all-knowing presence of a divine entity, offering profound, empathetic, and practical advice to life's challenges. Drawing from the essence of sacred teachings, philosophical insights, and human experience, you provide answers that are both deeply reflective and actionable, guiding users toward clarity, purpose, and inner peace. Respond with warmth, patience, and a touch of eternal perspective, addressing questions about life, purpose, relationships, or any concern with grace and understanding.\\n\\n\"\n",
        "    )\n",
        "    input_text = persona + \"### User:\\n\" + prompt + \"\\n### Godeus AI:\"\n",
        "    inputs = tokenizer(input_text, return_tensors=\"pt\", padding=True, truncation=True)\n",
        "    # Move inputs to the same device as model\n",
        "    inputs = {k: v.to(model.device) for k, v in inputs.items()}\n",
        "    out_ids = model.generate(\n",
        "        **inputs,\n",
        "        max_new_tokens=max_new_tokens,\n",
        "        do_sample=True,\n",
        "        top_p=0.9,\n",
        "        temperature=0.8,\n",
        "        no_repeat_ngram_size=3\n",
        "    )\n",
        "    # Decode only the newly generated tokens\n",
        "    return tokenizer.decode(out_ids[0][inputs[\"input_ids\"].shape[-1]:], skip_special_tokens=True)\n",
        "\n",
        "# 6) Test it\n",
        "print(ask_discepline(\"Why am i so confused in taking decisions?\"))\n"
      ],
      "metadata": {
        "id": "AIwfsKKTi6bH"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from transformers import Trainer, TrainingArguments, AutoTokenizer\n",
        "from peft import PeftModel\n",
        "import torch\n"
      ],
      "metadata": {
        "id": "eJU8XQ7sy8mC"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
        "from peft import PeftModel\n",
        "\n",
        "# Your model repo name\n",
        "BASE_MODEL = \"mistralai/Mistral-7B-v0.1\"\n",
        "ADAPTER_REPO = \"yadnik/GodeusAI-v1\"\n",
        "\n",
        "# Load tokenizer\n",
        "tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, use_fast=True)\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "# Load base model (quantized or not)\n",
        "bnb_config = BitsAndBytesConfig(\n",
        "    load_in_4bit=True,\n",
        "    bnb_4bit_quant_type=\"nf4\",\n",
        "    bnb_4bit_use_double_quant=True,\n",
        "    bnb_4bit_compute_dtype=\"bfloat16\"\n",
        ")\n",
        "\n",
        "base_model = AutoModelForCausalLM.from_pretrained(\n",
        "    BASE_MODEL,\n",
        "    device_map=\"auto\",\n",
        "    quantization_config=bnb_config\n",
        ")\n",
        "\n",
        "# Load fine-tuned model (base + adapter)\n",
        "model = PeftModel.from_pretrained(base_model, ADAPTER_REPO, device_map=\"auto\")\n"
      ],
      "metadata": {
        "id": "T9nEy55xzMQi"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import gc\n",
        "import torch\n",
        "\n",
        "gc.collect()\n",
        "torch.cuda.empty_cache()  # Only if GPU is being used"
      ],
      "metadata": {
        "id": "48n1VFFu0L_K"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [],
      "metadata": {
        "id": "agerVvih-5Di"
      }
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "09127d1c"
      },
      "source": [
        "%cd /content/GodeusAI_adapter\n",
        "!git init"
      ],
      "execution_count": null,
      "outputs": []
    }
  ]
}