Model Card for Mistral-QLoRA-Alpaca
Model Details
Model Description
This model is a QLoRA fine-tuned version of Mistral-7B trained on the Alpaca dataset for instruction-following tasks. It demonstrates parameter-efficient fine-tuning using LoRA adapters with 4-bit quantization.
- Developed by: Sujith Reddy
- Model type: Instruction-tuned causal language model (QLoRA adapter)
- Language: English
- License: Apache 2.0
- Finetuned from model: mistralai/Mistral-7B-v0.1
Model Sources
Uses
Direct Use
- Instruction following
- Question answering
- Educational and research purposes
- NLP experimentation
Downstream Use
- Can be integrated into chatbots
- Fine-tuned further for domain-specific tasks
Out-of-Scope Use
- Medical, legal, or financial advice
- Safety-critical applications
- Production systems without further validation
Bias, Risks, and Limitations
- Inherits biases from base model and training data
- May generate incorrect or misleading outputs
- Limited generalization due to small dataset (~5000 samples)
Recommendations
Users should validate outputs before using in real-world applications and avoid using in sensitive domains.
How to Get Started with the Model
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = "mistralai/Mistral-7B-v0.1"
adapter = "Sujith2121/mistral-qlora-alpaca"
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
prompt = "Explain machine learning in simple terms"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
Training Data
- Dataset: tatsu-lab/alpaca
- Type: Instruction-response dataset
- Samples used: ~5000
Training Procedure
Preprocessing
- Converted instruction, input, and output into formatted prompts
- Tokenized with max length of 256
Training Hyperparameters
- Training regime: fp16 mixed precision (QLoRA)
- Epochs: 1
- Learning rate: 2e-4
- Batch size: 2 with gradient accumulation
Speeds, Sizes, Times
- Training time: ~2–4 hours
- GPU: NVIDIA T4 (Kaggle)
Evaluation
Testing Data, Factors & Metrics
Testing Data
- Held-out subset (~50–100 samples) from Alpaca dataset
Factors
- Instruction-following quality
- Response completeness
Metrics
Results
Metric Base Model QLoRA Model
BLEU 18.5 27.2
ROUGE-L 30.1 42.8
Summary
The QLoRA model shows improved alignment with instructions and generates more detailed and relevant responses compared to the base model.
Environmental Impact
- Hardware Type: NVIDIA T4 GPU
- Hours used: ~3 hours
- Cloud Provider: Kaggle
- Compute Region: Unknown
- Carbon Emitted: Not calculated
Technical Specifications
Model Architecture and Objective
- Transformer-based architecture (Mistral-7B)
- Objective: Next-token prediction (causal language modeling)
Compute Infrastructure
Hardware
Software
- Transformers
- PEFT
- BitsAndBytes
- Datasets
- Evaluate
Citation
BibTeX:
@misc{mistral_qlora_alpaca,
author = {Sujith Reddy},
title = {Mistral QLoRA Alpaca},
year = {2026},
publisher = {Hugging Face}
}
APA:
Sujith Reddy. (2026). Mistral QLoRA Alpaca. Hugging Face.
Model Card Contact
Sujith Reddy
Framework versions
