Modelpublic

pmistryds/Maya-LLM-v1.0-bm

sourceHugging Faceapache-2.0updated 10mo agoView on Hugging Face

0likes8downloads

Model Card

Maya LLM v1.0

A LoRA Adapter for Mistral-7B-Instruct-v0.2 | Fine-tuned 7.25B Parameter Language Model

📖 Model Description

Technical Classification: LoRA (Low-Rank Adaptation) adapter for parameter-efficient fine-tuning

Maya is a LoRA adapter trained on Mistral-7B-Instruct-v0.2. When loaded with the base model, it creates a fine-tuned 7.25 billion parameter language model specialized for conversational AI tasks.

Dataset: 15,000 instruction-response pairs (13,500 training / 1,500 validation)

The adapter was trained on carefully curated examples, enabling expertise in:

—Python Programming (basic to advanced)
—Machine Learning & AI concepts
—Data Ethics & AI Guidelines
—Business Analysis
—Market Research
—SQL & Databases

Created by: Parth B Mistry Architecture: LoRA Adapter + Mistral-7B-Instruct-v0.2 Base Model: mistralai/Mistral-7B-Instruct-v0.2 Training Method: Supervised Fine-Tuning (SFT) with LoRA Model Type: Parameter-Efficient Fine-Tuned Language Model

⚡ Quick Start

Installation

Required Libraries:

bash

pip install transformers peft torch accelerate bitsandbytes

Usage

Note: This is a LoRA adapter, not a standalone model. You must load it with the base Mistral-7B model.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-Instruct-v0.2",
    device_map="auto",
    torch_dtype=torch.float16
)

# Load Maya adapter
model = PeftModel.from_pretrained(base_model, "pmistryds/Maya-LLM-v1.0-bm")
tokenizer = AutoTokenizer.from_pretrained("pmistryds/Maya-LLM-v1.0-bm")

# Generate response
prompt = "[INST] Who created you? [/INST]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Expected Output:

I was created by Parth B Mistry.

🎯 Model Details

Attribute	Value
Base Model	Mistral-7B-Instruct-v0.2
Total Parameters	7.25 Billion
Trainable Parameters	8.4 Million (LoRA)
Training Efficiency	0.116% trainable
Adapter Size	~50-100 MB
Full Model Size (with base)	~14 GB (base) + 50MB (adapter)
Dataset Size	15,000 instruction-response pairs
Training Split	13,500 examples (90%)
Validation Split	1,500 examples (10%)
Training Steps	800 steps
Training Time	~84 minutes on RTX A40
Final Validation Loss	1.206
Method	LoRA (r=16, alpha=32)

💡 Use Cases

Maya LLM excels at:

1.Code Assistance - Python programming help, debugging, best practices
2.ML/AI Education - Explaining machine learning concepts clearly
3.Data Ethics - Guidance on GDPR, data privacy, AI ethics
4.Business Analysis - Market research, data analysis insights
5.Database Queries - SQL help and optimization

🔧 Training Details

Training Configuration

—LoRA Rank: 16
—LoRA Alpha: 32
—LoRA Dropout: 0.05
—Target Modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
—Learning Rate: 5e-5 (conservative)
—Batch Size: 16 (effective)
—Max Steps: 800
—Optimizer: pagedadamw8bit
—Precision: bfloat16
—GPU: NVIDIA RTX A40 (48GB)

Training Results

The model was trained with early stopping monitoring to prevent overfitting:

Step	Training Loss	Validation Loss	Accuracy
100	1.664	1.623	63.1%
200	1.330	1.353	66.6%
300	1.264	1.287	67.9%
400	1.204	1.251	68.6%
500	1.170	1.228	69.0%	← Best
600	1.182	1.212	69.3%
700	1.189	1.206	69.4%
800	1.179	1.206	69.4%

Best checkpoint: Step 500 (automatically selected)

🚀 Deployment Options

Option 1: Gradio UI

python

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
model = PeftModel.from_pretrained(base, "pmistryds/Maya-LLM-v1.0-bm")
tokenizer = AutoTokenizer.from_pretrained("pmistryds/Maya-LLM-v1.0-bm")

def chat(message, history):
    prompt = f"[INST] {message} [/INST]"
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=200)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split("[/INST]")[1].strip()

gr.ChatInterface(fn=chat, title="Maya LLM").launch()

Option 2: FastAPI Server

python

from fastapi import FastAPI
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

app = FastAPI()
base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
model = PeftModel.from_pretrained(base, "pmistryds/Maya-LLM-v1.0-bm")
tokenizer = AutoTokenizer.from_pretrained("pmistryds/Maya-LLM-v1.0-bm")

@app.post("/chat")
def chat(message: str):
    prompt = f"[INST] {message} [/INST]"
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=200)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"response": response.split("[/INST]")[1].strip()}

📊 Example Interactions

Example 1: Identity

User: Who created you?
Maya: I was created by Parth B Mistry.

Example 2: Basic Math

User: What is 2+2?
Maya: 2+2 equals 4.

Example 3: ML Explanation

User: Explain machine learning briefly.
Maya: Machine learning is a type of AI where a model learns patterns 
from data instead of being explicitly programmed. It improves over 
time as it sees more examples.

⚠️ Limitations

—Based on Mistral-7B (early 2024 training cutoff)
—May not have information on very recent events
—Best suited for educational and conversational purposes
—Adapter-based approach requires base model download

📝 Citation

bibtex

@misc{maya-llm-v1,
  author = {Parth B Mistry},
  title = {Maya LLM: A Personalized AI Assistant},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/pmistryds/Maya-LLM-v1.0-bm}},
}

🔗 Links

—Creator GitHub: ParthDS02
—Model Repository: pmistryds/Maya-LLM-v1.0-bm
—Base Model: Mistral-7B-Instruct-v0.2
—Training Framework: TRL by HuggingFace

📜 License

This model inherits the Apache 2.0 license from the base Mistral-7B-Instruct-v0.2 model.

🙏 Acknowledgments

—Mistral AI for the excellent base model
—HuggingFace for the training libraries (Transformers, TRL, PEFT)
—Community for LoRA and efficient fine-tuning techniques

Built with ❤️ by Parth B Mistry