Fine-tuned a Large Language Model on NLP and RAG domain knowledge using QLoRA — training only 0.089% of parameters on a free GPU.
Published Model: huggingface.co/Rohith2026/nlp-rag-expert
A domain-specific fine-tuned LLM that specialises in answering NLP and RAG questions — my exact area of expertise from my MSc thesis on hybrid retrieval.
The model was fine-tuned using QLoRA (Quantised Low-Rank Adaptation) — the industry-standard technique for efficient LLM fine-tuning used at every serious AI company.
Standard fine-tuning requires updating 100% of model parameters. For a 1.1B model that means storing and updating 1.1 billion values — impossible on a free 15GB T4 GPU.
QLoRA solves this elegantly:
Full model (frozen, 4-bit) ──► Only 0.089% of parameters trained
↓ ↓
4x memory reduction Small LoRA matrices added
↓ ↓
Fits on free T4 GPU Comparable to full fine-tuning
This is the same technique used to fine-tune 7B to 70B models in production.
| Metric | Value |
|---|---|
| Base model | TinyLlama 1.1B Chat |
| Fine-tuning method | QLoRA (4-bit NF4 quantisation) |
| LoRA rank | 16 |
| Total parameters | 1.1 Billion |
| Parameters trained | 0.089% only |
| Training examples | 10 NLP/RAG instruction pairs |
| Starting loss | 2.47 |
| Final loss | 0.89 |
| Epochs | 3 |
| Training time | Under 10 minutes |
| Hardware | NVIDIA T4 GPU (Google Colab free tier) |
| Tool | Purpose | Why I chose it |
|---|---|---|
| Unsloth | QLoRA training library | 2x faster + 60% less memory than standard HuggingFace |
| HuggingFace PEFT | LoRA adapter management | Industry standard — in 70% of ML job descriptions |
| TRL SFTTrainer | Supervised fine-tuning | Used at HuggingFace and major AI labs |
| bitsandbytes | 4-bit quantisation | Essential for running LLMs on limited hardware |
| HuggingFace Hub | Model publishing | Public proof of fine-tuning skills |
| Google Colab T4 | Free GPU | Accessible to anyone — no cloud spend needed |
The model was fine-tuned to answer expert questions about:
- Retrieval-Augmented Generation (RAG) systems
- BM25 vs dense retrieval — differences and tradeoffs
- FAISS vector databases and index types
- LoRA and QLoRA fine-tuning techniques
- LLM evaluation frameworks (RAGAS, MRR, Recall@K)
- Semantic search and dense embeddings
- Hybrid retrieval strategies
This domain was chosen because it is my thesis area — fine-tuning on what you know deeply produces better training examples.
With 10 training examples on a 1.1B model, the experiment demonstrated the complete QLoRA workflow end-to-end. The model showed partial domain adaptation — the technique works, but scale matters.
What I learned:
- Larger datasets (1000+ examples) produce much stronger domain specialisation
- Bigger models (7B+) respond better to fine-tuning
- Free Colab T4 constraints limited this experiment intentionally
- The workflow is identical at scale — only the dataset size and model change
This is the honest engineering insight: understanding the limits of an approach is as valuable as making it work.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(
"TinyLlama/TinyLlama-1.1B-Chat-v1.0"
)
model = PeftModel.from_pretrained(base, "Rohith2026/nlp-rag-expert")
tokenizer = AutoTokenizer.from_pretrained("Rohith2026/nlp-rag-expert")
prompt = "### Instruction:\nWhat is RAG?\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
| Project | Description | Live |
|---|---|---|
| Hybrid RAG System | BM25 + dense embeddings · 93% Recall@10 · 8.84M passages · MSc thesis | Demo · API |
| AI Agent System | ReAct agent · LangGraph · 3 tools · web search + calculator + RAG | Demo |
| LLM Fine-Tuning (this) | QLoRA · TinyLlama 1.1B · NLP/RAG domain · published model | Model |
Rohith Kumar Reddipogula MSc Data Science — University of Europe for Applied Sciences, Berlin
