README.md

---
base_model: unsloth/qwen2.5-14b-instruct-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
- trl
license: apache-2.0
language:
- en
---

# Uploaded  model

- **Developed by:** Marcos Gôlo
- **License:** apache-2.0
- **Finetuned from model :** unsloth/qwen2.5-14b-instruct-bnb-4bit

This qwen2 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

## 📄 Model Card: `aksw/text2sparql-M`

### 🧠 Model Overview

`text2sparql-M` is a Medium fine-tuned language model designed to **translate natural language questions into SPARQL queries**, specifically targeting the **DBpedia knowledge graph (2014 version)**. It is ideal for knowledge-based QA systems and symbolic reasoning agents.

---

### 🔍 Intended Use

* **Input**: Natural language questions (e.g., *"Which actors were born in Germany?"*)
* **Output**: A **single string** containing the corresponding SPARQL query.

---

### 🧩 Applications

* Question Answering systems over open knowledge bases (DBpedia)
* Semantic conversational agents
* Knowledge graph exploration tools
* Autonomous agents with symbolic reasoning capabilities

---

### ⚙️ Model Details

* **Base model**: Qwen2.5 14B (via Unsloth)
* **Training**: Dataset with 15.000 question-query examples built by joining 4 datasets:
  * QLAD-1
  * LCQUAD-1
  * ParaQA 
  * Question-Sparql
* **Target Ontology**: DBpedia Ontology (2014)
* **Frameworks**: Unsloth, HuggingFace, Transformers

---

### 📦 Installation

Make sure to install `unsloth`, `torch` and CUDA dependencies:

```bash
pip install unsloth torch
```

---

### 🧪 Example: Inference Code

```python
from unsloth import FastLanguageModel
import torch

class SPARQLQueryGenerator:
    def __init__(self, model_name: str, max_seq_length: int = 2048, load_in_4bit: bool = True):
        self.model, self.tokenizer = FastLanguageModel.from_pretrained(
            model_name=model_name,
            max_seq_length=max_seq_length,
            load_in_4bit=load_in_4bit
        )
        _ = FastLanguageModel.for_inference(self.model)

    def build_prompt(self, question: str) -> list:
        return [
            {"role": "system", "content": (
                "You are an expert data analyst with deep knowledge of SPARQL and the DBpedia ontology.\n"
                "Your task is to convert a given natural language question into a syntactically correct DBpedia SPARQL query "
                "that accurately retrieves the answer.\n"
                "Your output must be a single string containing only the SPARQL query—no additional text, explanation, or commentary.\n"
                "Ensure that you use the appropriate DBpedia prefixes and follow standard SPARQL syntax."
            )},
            {"role": "user", "content": question}
        ]

    def generate_query(self, question: str, temperature: float = 0.01, max_new_tokens: int = 1024) -> str:
        messages = self.build_prompt(question)
        inputs = self.tokenizer.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=True,
            return_tensors="pt"
        ).to("cuda")

        outputs = self.model.generate(
            input_ids=inputs,
            max_new_tokens=max_new_tokens,
            use_cache=True,
            temperature=temperature,
            min_p=0.1
        )

        decoded = self.tokenizer.batch_decode(outputs)[0]
        return self._extract_sparql(decoded)

    def _extract_sparql(self, decoded_text: str) -> str:
        start_token = "<|im_start|>assistant\n"
        end_token = "<|im_end|>"
        start_index = decoded_text.find(start_token) + len(start_token)
        sparql = decoded_text[start_index:]
        return sparql.rstrip(end_token) if sparql.endswith(end_token) else sparql

# --- Using the model ---
if __name__ == "__main__":
    generator = SPARQLQueryGenerator(model_name="aksw/text2sparql-M")
    question = "Which actors were born in Germany?"
    query = generator.generate_query(question)
    print(query)
```

---

### 🧠 Example Input / Output

#### Input:

```text
Which actors were born in Germany?
```

#### Output:

```sparql
PREFIX dbo: <http://dbpedia.org/ontology/>
PREFIX res: <http://dbpedia.org/resource/>
PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
SELECT DISTINCT ?uri WHERE {
    ?uri rdf:type dbo:Actor .
    ?uri dbo:birthPlace res:Germany .
}
```

---

### 🧪 Evaluation

The model was evaluated using F1-score on a hand-crafted dataset for the First Text2SPARQL Challenge, a Co-Located with Text2KG at ESWC25.

---

### 📚 Citation

If you use this model in your work, please cite it as:

```
@misc{text2sparql2025,
  author = {Marcos Gôlo, Paulo do Carmo, Edgard Marx, Ricardo Marcacini},
  title = {Text2SPARQL-M: Natural Language Text to SPARQL for DBpedia},
  year = {2025},
  howpublished = {\url{https://huggingface.co/aksw/text2sparql-M}},
}
```