|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- pt |
|
|
base_model: |
|
|
- google/gemma-3-270m |
|
|
--- |
|
|
|
|
|
# 🐶 DogeAI-v1.5-Coder |
|
|
|
|
|
DogeAI-v1.5-Coder is a **small, experimental code-focused language model** fine-tuned from **Gemma 3 (270M parameters)**. |
|
|
|
|
|
This model was created as a learning and experimentation project, focusing on **code generation and completion** with limited resources. It is **not intended to compete with large-scale coding models**, but rather to explore how far a compact model can go when domain-focused. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🔍 Model Details |
|
|
|
|
|
- **Base model:** Gemma 3 – 270M |
|
|
- **Fine-tuning type:** Supervised fine-tuning (SFT) |
|
|
- **Primary domain:** Programming / code-related text |
|
|
- **Languages:** Mixed (depends on dataset; mainly scripting-style code) |
|
|
- **Parameters:** ~270 million |
|
|
- **Context length:** Limited (inherits base model constraints) |
|
|
|
|
|
--- |
|
|
|
|
|
## 🎯 Intended Use |
|
|
|
|
|
DogeAI-v1.5-Coder is best suited for: |
|
|
|
|
|
- Simple code completion |
|
|
- Small scripting examples |
|
|
- Educational purposes (learning how fine-tuning works) |
|
|
- Research on **small language models** |
|
|
- Benchmarking and experimentation |
|
|
|
|
|
It performs best when: |
|
|
- Prompts are short and explicit |
|
|
- The task is narrow and well-defined |
|
|
- Expectations are aligned with its size |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚠️ Limitations |
|
|
|
|
|
This model has **clear and expected limitations**: |
|
|
|
|
|
- Weak long-range reasoning |
|
|
- Inconsistent performance on complex programming tasks |
|
|
- Limited generalization outside the training distribution |
|
|
- Not reliable for production or critical systems |
|
|
|
|
|
These limitations are a direct consequence of its **small scale and experimental nature**. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧪 Training Notes |
|
|
|
|
|
- The model was fine-tuned on a custom dataset focused on code-related text. |
|
|
- No reinforcement learning or advanced alignment techniques were used. |
|
|
- The goal was experimentation and learning, not optimization for benchmarks. |
|
|
|
|
|
--- |
|
|
|
|
|
## 📚 Why This Model Exists |
|
|
|
|
|
DogeAI-v1.5-Coder exists as a **learning artifact**. |
|
|
|
|
|
It represents: |
|
|
- Early experimentation with fine-tuning |
|
|
- Exploration of low-parameter models |
|
|
- A step in understanding data quality, formatting, and model behavior |
|
|
|
|
|
Small models are valuable tools for understanding how language models actually work. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚫 What This Model Is NOT |
|
|
|
|
|
- ❌ A replacement for large coding assistants |
|
|
- ❌ A reasoning-focused model |
|
|
- ❌ Production-ready |
|
|
- ❌ Instruction-following at a high level |
|
|
|
|
|
--- |
|
|
|
|
|
## 📜 License |
|
|
|
|
|
This model follows the same license as its base model (Gemma). |
|
|
Please ensure compliance with the original license when using or redistributing. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🙌 Acknowledgements |
|
|
|
|
|
- Google Gemma team for the base model |
|
|
- The open-source ML community |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧠 Final Note |
|
|
|
|
|
DogeAI-v1.5-Coder is small, imperfect, and honest. |
|
|
Its value lies in experimentation, not performance. |
|
|
|
|
|
Sometimes, understanding the limits teaches more than chasing scale. |
|
|
|
|
|
MADE BY AXIONLAB |
