README.md

---
license: apache-2.0
language: en
library_name: transformers
pipeline_tag: text-generation
datasets:
  - Open-Orca/OpenOrca
  - Open-Orca/SlimOrca
base_model: HuggingFaceTB/SmolLM3-3B
tags:
  - qlora
  - smollm3
  - fine-tuned
  - rag
---

# 🧠 SmolLM3 QLoRA - OpenOrca Fine-Tuned

**SmolLM3 QLoRA** is a lightweight, 3B parameter open-source language model based on [SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B), fine-tuned using [QLoRA](https://arxiv.org/abs/2305.14314) on the [OpenOrca Slim](https://huggingface.co/datasets/Open-Orca/SlimOrca) dataset (500K examples). It is optimized for **retrieval-augmented generation (RAG)** use cases and delivers **competitive benchmark scores** against much larger models like LLaMA-2 7B.

---

## ✨ Model Highlights

- 🔍 **Trained for real-world queries** using OpenOrca-style assistant data.
- ⚡ **Efficient:** 3B parameter model that runs on a single A100 or consumer GPU.
- 🧠 **Competent generalist:** Performs well on reasoning and knowledge tasks.
- 🔗 **RAG-friendly:** Ideal for hybrid search setups using BM25 + FAISS.
- 🧪 **Evaluated on benchmarks:** Outperforms similar-sized models.

---

## 🧰 Intended Use

SmolLM3 QLoRA is intended to serve as a fast and compact assistant model for:

- 💬 Lightweight RAG pipelines
- 📚 Document and web snippet reasoning
- 🤖 Prototype assistants
- 🧪 AI research in instruction tuning and hybrid retrieval

---

## 🧪 Evaluation

The model has been evaluated using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) on a 500-sample subset of key academic benchmarks.

| Task           | Accuracy | Normalized Accuracy | LLaMA-2 7B |
|----------------|----------|---------------------|------------|
| **HellaSwag**  | 51.2%    | 66.4%               | 56.7% / 73.2% |
| **ARC-Challenge** | 49.4% | 52.2%               | 53.7% / 56.9% |
| **BoolQ**      | 81.0%    | —                   | 83.1% |

👉 Model achieves **~90–95% of LLaMA-2 7B** performance at less than **half the size**.

---

## 🏗️ Training Configuration

- **Base Model:** [`SmolLM3-3B`](https://huggingface.co/HuggingFaceTB/SmolLM3-3B)
- **Finetuning Method:** QLoRA (LoRA rank=8)
- **Dataset:** `Open-Orca/SlimOrca` (500K samples)
- **Precision:** bfloat16
- **Epochs:** 3
- **Max Length:** 1024 tokens
- **Hardware:** 2x A100 80GB
- **Framework:** 🤗 Transformers + TRL

---

## 🧠 How to Use

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "soupstick/smollm3-qlora-ft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

inputs = tokenizer("Explain retrieval-augmented generation.", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))