| --- |
| language: |
| - en |
| pipeline_tag: text-generation |
| license: apache-2.0 |
| base_model: Qwen/Qwen2.5-0.5B-Instruct |
| tags: |
| - unsloth |
| - qwen |
| - text-generation |
| - code |
| --- |
| |
| # Model Card for GEAR-2-500m-Identity |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| GEAR-2-500m-Identity is a lightweight Transformer LLM with approximately 0.5 billion parameters, fine-tuned on the Qwen2.5 architecture using Unsloth. It is designed to run extremely fast on local machines (CPU/Edge) with minimal memory usage. The model embodies the persona of **Gear**, an intelligent assistant created by **HeavensHack**. |
|
|
| It is capable of code generation (Python) and general chat. While efficient, it is a small model and may struggle with complex reasoning compared to larger parameters. |
|
|
| - **Developed by:** HeavensHack |
| - **Model type:** Qwen2 For Causal LM |
| - **Language(s) (NLP):** English, Python (Code) , **(New)** *Russian* |
| - **License:** Apache 2.0 |
| - **Finetuned from model:** Qwen/Qwen2.5-0.5B-Instruct |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| - Fast local chat assistant |
| - Python code generation and debugging |
|
|
| ### Out-of-Scope Use |
|
|
| - Complex mathematical reasoning |
| - High-stakes decision making |
| - Long-context analysis requiring high accuracy |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - **Hallucinations:** Due to the 0.5B parameter size, it may generate plausible but incorrect information. |
| - **Identity:** The model is strictly fine-tuned to identify itself as "Gear" by HeavensHack. |
| - **Inconsistency:** Behavior might be variable in long conversations. |
|
|
| ### Recommendations |
|
|
| - Use for educational purposes, hobby projects, or low-resource environments. |
| - Verify any code generated before running it in production. |
|
|
| ## How to Get Started |
|
|
| - Load the model using Unsloth or standard Hugging Face transformers. |
| - Optimized for local inference. |
|
|
| ## Training Details |
|
|
| - **Training Data:** Custom identity dataset (HeavensHack), Alpaca (English), and Python Code instructions. |
| - **Training Procedure:** Fine-tuned using Unsloth (LoRA) for efficiency. |
| - **Training Regime:** Mixed precision (BF16/FP16). |
|
|
| ## Evaluation |
|
|
| - Validated for identity retention and basic coding tasks. |
| - Not benchmarked for enterprise production use. |
|
|
| ## Environmental Impact |
|
|
| - Extremely low compute cost during training due to Unsloth optimization. |
|
|
| ## Model Card Contact |
|
|
| - **Author:** HeavensHackDev |
|
|
| # But... |
|
|
| - **At first, only the GGUF file will be available. The rest will follow later.** |
