metadata
license: mit
tags:
- tinyllama
- sciq
- multiple-choice
- peft
- lora
- 4bit
- quantization
- instruction-tuning
datasets:
- allenai/sciq
language:
- en
library_name: transformers
pipeline_tag: text-generation
🧠 TinyLLaMA-1.1B LoRA Fine-tuned on SciQ Dataset
This is a TinyLLaMA-1.1B model fine-tuned using LoRA (Low-Rank Adaptation) on the SciQ multiple-choice question answering dataset. It uses 4-bit quantization via bitsandbytes to reduce memory usage and improve inference efficiency.
🧪 Use Cases
This model is suitable for:
- Educational QA bots
- MCQ-style reasoning
- Lightweight inference on constrained hardware (e.g., GPUs with <8GB VRAM)
🛠️ Training Details
- Base Model:
TinyLlama/TinyLlama-1.1B-Chat-v1.0 - Dataset:
allenai/sciq(Science QA) - Method: Parameter-Efficient Fine-Tuning using LoRA
- Quantization: 4-bit using
bitsandbytes - Framework: 🤗 Transformers + PEFT + Datasets
🧬 Model Architecture
- Model: Causal Language Model
- Fine-tuned layers:
q_proj,v_proj(via LoRA) - Quantization: 4-bit (bnb config)
📊 Evaluation
- Accuracy: 100% on a 1000-sample SciQ subset
- Eval Loss: ~0.19
💡 How to Use
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("your-username/tinyllama-sciq-lora")
tokenizer = AutoTokenizer.from_pretrained("your-username/tinyllama-sciq-lora")
prompt = """Question: What is the boiling point of water?\nChoices:\nA. 50°C\nB. 75°C\nC. 90°C\nD. 100°C\nAnswer:"""
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
## 🔐 License
This model is released under the MIT License.
## 🙌 Credits
FineTuned By - [Uditanshu Pandey](https://huggingface.co/TechyCode)
Based on - [TinyLLaMA-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0)