| # light_gpt_text_generator | |
| ## Overview | |
| `light_gpt_text_generator` is a distilled, 6-layer generative Transformer based on the GPT-2 architecture. It is optimized for low-latency text completion and creative writing tasks where computational resources are limited (Edge devices/Mobile). | |
| ## Model Architecture | |
| - **Type**: Causal Language Model (Decoder-only) | |
| - **Layers**: 6 Transformer blocks (Reduced from GPT-2 Base's 12). | |
| - **Embedding Dim**: 768 | |
| - **Heads**: 12 Multi-head self-attention. | |
| - **Tokenizer**: Byte-level Byte Pair Encoding (BPE). | |
| ## Intended Use | |
| - Real-time autocomplete for code or prose. | |
| - Creative writing assistance and brainstorming. | |
| - Chatbot prototyping for specific domains. | |
| ## Limitations | |
| - **Hallucination**: High tendency to generate factually incorrect information. | |
| - **Coherence**: Difficulty maintaining logical consistency over very long passages (>500 words). | |
| - **Safety**: Lack of an integrated RLHF layer means the model may generate toxic or biased content if prompted inappropriately. |