Buckets:
| pipeline_tag: text-generation | |
| library_name: pytorch bitsandbytes | |
| datasets: | |
| - AxisCommunity/Dataset_ModelDucky | |
| language: | |
| - ru | |
| - en | |
| - kk | |
| # Model Card for DuckyModel_V1 | |
| LLaMA-style decoder-only transformer, обученная с нуля. | |
| ## Model Details | |
| ### Model Description | |
| Языковая модель на ~1.05B параметров, архитектура в стиле LLaMA (RMSNorm, RoPE, SwiGLU, Flash Attention, weight tying). Обучена с нуля на собственном BPE-токенизаторе. | |
| - **Developed by:** AxisCommunity | |
| - **Model type:** Decoder-only transformer (text generation) | |
| - **Language(s):** Russian, English, Kazakh | |
| - **License:** [More Information Needed] | |
| ### Model Sources | |
| - **Repository:** https://huggingface.co/AxisCommunity/DuckyModel_V1 | |
| ## Uses | |
| ### Direct Use | |
| Генерация текста на русском, английском, казахском языках. | |
| ### Out-of-Scope Use | |
| Модель обучена на ограниченном объёме данных (300 шагов), не предназначена для высокоточных или критичных задач. | |
| ## Bias, Risks, and Limitations | |
| Модель обучена на небольшом количестве шагов и может выдавать несвязный или некорректный текст. | |
| ## How to Get Started with the Model | |
| Модель использует кастомную архитектуру (не из transformers), весовой файл в формате safetensors с 4-bit квантованием. | |
| ## Training Details | |
| ### Training Data | |
| Датасет: AxisCommunity/Dataset_ModelDucky (подвыборка ~20000 примеров) | |
| ### Training Procedure | |
| #### Training Hyperparameters | |
| - **Training regime:** bf16 mixed precision, AdamW8bit optimizer | |
| - **Steps:** 300 | |
| - **Batch size:** 2 | |
| - **Learning rate:** 3e-4 | |
| - **Sequence length:** 512 | |
| #### Speeds, Sizes, Times | |
| - **Размер модели:** ~600MB (4-bit quantized) | |
| ## Technical Specifications | |
| ### Model Architecture and Objective | |
| - Параметров: ~1.05B | |
| - Слоёв: 20 | |
| - Hidden size: 2048 | |
| - Attention heads: 16 | |
| - Intermediate size (FFN): 5632 | |
| - Vocab size: 9482 | |
| - Компоненты: RMSNorm, RoPE, SwiGLU, Flash Attention (scaled_dot_product_attention), weight tying | |
| ### Compute Infrastructure | |
| #### Software | |
| PyTorch, bitsandbytes, tokenizers, safetensors |
Xet Storage Details
- Size:
- 2.45 kB
- Xet hash:
- 2861bcf74b0f42b944c66609470adbbdeaa7c24e88100f5fd3dee6d36555b4e0
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.