79 lines
2.8 KiB
Markdown
79 lines
2.8 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen2.5-3B
|
||
|
|
tags:
|
||
|
|
- code
|
||
|
|
- text-generation
|
||
|
|
- code-assistant
|
||
|
|
- csharp
|
||
|
|
- sql
|
||
|
|
- react
|
||
|
|
- project-lsda
|
||
|
|
- synthetic-data
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- tr
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# LSDA-3B-Turkish-Dev
|
||
|
|
|
||
|
|
This model is a high-performance LLM specifically trained for modern full-stack development with a deep focus on **C#**, **SQL**, and **React**.
|
||
|
|
|
||
|
|
## 💎 Dataset & Methodology
|
||
|
|
Unlike many small-scale models that rely on raw web crawls, **LSDA-3B-Turkish-Dev** was trained using a **Curated and Artificially Augmented dataset** specifically designed for full-stack workflows. This ensures high-quality weight updates and robust convergence for complex coding patterns.
|
||
|
|
|
||
|
|
- **C# & React Synergy:** The dataset includes thousands of cross-referenced examples between backend APIs and frontend components.
|
||
|
|
- **SQL Precision:** Augmented query-schema pairs to improve complex join logic.
|
||
|
|
- **LSDA Framework:** Our proprietary augmentation process ensures high-quality weight updates and robust convergence, even for complex coding patterns.
|
||
|
|
- **Bilingual Logic:** Engineered to maintain high coding standards while providing fluent technical explanations in both English and Turkish.
|
||
|
|
|
||
|
|
## 🎯 Specialized Domains (The Big Three)
|
||
|
|
|
||
|
|
The model is heavily optimized for:
|
||
|
|
|
||
|
|
- **C# & .NET:** Professional backend architecture, LINQ, and modern .NET patterns.
|
||
|
|
- **SQL:** High-level query generation, optimization, and DDL/DML tasks.
|
||
|
|
- **React:** Component lifecycle, state management (Hooks/Context), and modern UI logic.
|
||
|
|
|
||
|
|
## 🌐 Language Support
|
||
|
|
|
||
|
|
Strictly optimized for a bilingual experience:
|
||
|
|
1. **English:** Global software engineering standards.
|
||
|
|
2. **Turkish:** Fully localized technical explanations and Turkish documentation support.
|
||
|
|
|
||
|
|
*Note: It is highly recommended to use the model within these two languages for best results.*
|
||
|
|
|
||
|
|
|
||
|
|
## 🚀 Model Details
|
||
|
|
|
||
|
|
- **Type:** Full Model (Ready to use).
|
||
|
|
- **Architecture:** Qwen2.5
|
||
|
|
- **Training Env:** Optimized via LSDA Data Augmentation Framework.
|
||
|
|
- **Format:** Safetensors (Sharded)
|
||
|
|
|
||
|
|
|
||
|
|
## 💻 Usage Example
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model_id = "umitaksoylu/lsda-3b-turkish-dev"
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
|
|
|
||
|
|
# Example: Bridging C# and React
|
||
|
|
prompt = "Write a C# DTO class and a corresponding React interface for a User Profile."
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "You are a senior developer assistant. You are a helpful assistant for C#, SQL and React development."},
|
||
|
|
{"role": "user", "content": prompt}
|
||
|
|
]
|
||
|
|
|
||
|
|
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
|
|
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
|
|
|
||
|
|
outputs = model.generate(**inputs, max_new_tokens=1024)
|
||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|