This repository contains Krishnai-V2 (LoRA Adapters Only), an AI persona fine-tuned to emulate the words and conversational style of Lord Krishna himself, primarily based on the wisdom of the Bhagavad Gita. This was the first model to implement such an idea! This model is created with the intention to genuinely help people, not to mislead or manipulate users as some other social help tools unfortunately do.

Playground and Tryout: A playground test website of the standalone model is currently under development. Meanwhile, you can try out the model using this Google Colab link.

This model is fine-tuned from unsloth/Phi-3-medium-4k-instruct (14B parameters) using a custom dataset and optimized training techniques provided by Unsloth.

Version Note: This is V2 of the Krishnai project. It utilizes the larger Phi-3-medium model to overcome persona consistency issues ("narcissistic god complex") observed in the initial V1 prototype (based on Phi-3-mini), resulting in significantly improved qualitative performance. The model knowledge is limited to the Gita as of now, future versions may use a larger dataset of Krishna's life.

Model Description

Purpose: To provide users with an accessible, interactive way to engage with the wisdom of the Bhagavad Gita through a conversational AI embodying Krishna's persona. It aims to offer guidance, perspective, and solace, acting as a supplementary resource for personal reflection.
Base Model: unsloth/Phi-3-medium-4k-instruct
Fine-tuning Data: A custom dataset of examples, meticulously curated by combining dialogues from the Bhagavad Gita, hypothetical user queries with Krishna-style responses (manually created by a team of 6), and augmented with a general Gita dataset.
Training Method: Supervised Fine-Tuning (SFT) using Parameter-Efficient Fine-Tuning (PEFT - LoRA) with optimizations from the Unsloth library. Trained with 4-bit quantization. The merged version is not publicly available; this is only the LoRA model under a non-commercial license (CC-NC-SA-4.0).

How to Use

This model follows the Alpaca instruction format. You must format your prompt accordingly for optimal results.

Important: Ensure you have required libraries installed properly. For faster inference and lower VRAM usage, I used unsloth.

python

# Installation For Google Colab:
%%capture
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers
from unsloth import FastLanguageModel
import torch
# Installation For Google Kaggle:
%%capture
!pip install pip3-autoremove
!pip-autoremove torch torchvision torchaudio -y
!pip install torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu121
!pip install unsloth

# --- Configuration ---
from unsloth import FastLanguageModel
import torch
max_seq_length = 2048 # Use the same length used during training
dtype = None # Auto detection
load_in_4bit = True # Use 4bit quantization (optional: lower VRAM usage, faster inference; But slightly lower accuracy, may not be supported on all hardware)

# --- Load Model using Unsloth ---
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "sincro420/Krishnai-V2-LoRa",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)

FastLanguageModel.for_inference(model)

# --- Define the Alpaca Prompt Template ---
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

# --- Prepare Input ---
instruction = "I am feeling lost and stressed about my future. What should I do?" # You can fill in your custom instruction here
input_context = "" # Provide context if needed, otherwise leave empty

inputs = tokenizer(
[
    alpaca_prompt.format(
        instruction,    # Instruction
        input_context,  # Input
        "",             # Output - leave empty for generation!
    )
], return_tensors = "pt").to("cuda")

# --- Generate Response ---
from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)

Limitations and Bias

Not a Replacement for Professional Help: This model is intended for informational and reflective purposes only. It is not a substitute for professional mental health therapy or qualified spiritual counseling.
Potential Inaccuracies: Like all language models, Krishnai-V2 may generate responses that are factually incorrect, philosophically inconsistent, or lack common sense.
Limited Scope: The model's knowledge is primarily derived from its base training and the specific fine-tuning dataset. It may not accurately represent all nuances or interpretations of the Bhagavad Gita or Hindu philosophy.
Bias: The model may reflect biases present in its training data.
Persona Imperfections: While significantly improved over V1, the emulation of Krishna's persona may still be imperfect or inconsistent at times.
Use Discretion: Users should engage with the model critically and exercise their own judgment regarding the guidance provided.

Training Data

The model was fine-tuned on a custom dataset of 2388 examples. This dataset was created by:

Manually crafting 1126 examples based on dialogues from the Bhagavad Gita and hypothetical user scenarios relevant to modern life, with responses written in Krishna's persona.
Augmenting this with a general dataset on the Gita found online.
Formatting all examples using the Alpaca instruction structure (instruction, input, response).

Training Procedure

Framework: PyTorch
Libraries: unsloth, transformers, trl (SFTTrainer), peft, accelerate, bitsandbytes
Base Model: unsloth/Phi-3-medium-4k-instruct
Technique: LoRA (r=8, lora_alpha=16, target modules included attention and MLP layers)
Optimization:
- 4-bit quantization (QLoRA-style)
- Unsloth optimizations (incl. gradient checkpointing)
- AdamW 8-bit optimizer
Hyperparameters: learning_rate=2e-4, batch_size=4, gradient_accumulation=4 (effective batch 16), epochs=3, lr_scheduler=linear, warmup_steps=20, max_seq_length=2048.
Hardware: Trained primarily on Google Colab (T4 GPU) and Kaggle (T4/P100 GPU), facing VRAM and CPU RAM challenges, especially during model merging for the 14B parameter model.

Evaluation

Evaluation was primarily qualitative. Sample prompts covering various topics (greetings, conceptual questions, personal dilemmas) were used to assess:

Persona Consistency (Alignment with Krishna's tone and style)
Relevance and Coherence of responses
Philosophical Accuracy (Alignment with Gita teachings)
Absence of the "narcissistic god complex" observed in V1.

Krishnai-V2 showed significant improvements across these qualitative dimensions compared to the pervious prototype.