| --- |
| license: apache-2.0 |
| datasets: |
| - HuggingFaceFW/fineweb-edu |
| language: |
| - en |
| tags: |
| - small |
| - cpu |
| - fast |
| - opensource |
| - open |
| - free |
| base_model: |
| - LH-Tech-AI/Apex-1-Instruct-350M |
| pipeline_tag: text-generation |
| --- |
| |
| **Apex 1.5: Improved reasoning and logic. Fixed wrong facts and hallucinations by increasing FineWeb-Edu ratio while finetuning to 4:1.** |
|
|
| **Update**: Due to the great community feedback on Apex 1.0, I've trained Apex 1.5 with a focus on world knowledge (FineWeb-Edu) and coding logic. Enjoy the massive jump in reasoning! |
|
|
| # How to train it |
| You can train it, using the **base model** of LH-Tech-AI/Apex-1-Instruct-350M (you need to train that base model first!). |
| Then, use the prepare-script and the finetuning script in the files list of this HF model. |
|
|
| # How to use it |
| You can download the `apex_1.5.gguf` or use `ollama run hf.co/LH-Tech-AI/Apex-1.5-Instruct-350M`. And you can also use it in LM Studio for example, just by searching for "Apex 1.5". |
|
|
| You can directly download the final model as ONNX format - so it runs without the need to install a huge Python environment with PyTorch, CUDA, etc... - as INT8 and in full precision. |
| Use `inference.py` for local inference on CUDA or CPU! First, install `pip install onnxruntime-gpu tiktoken numpy nvidia-cudnn-cu12 nvidia-cublas-cu12` on your system (in a Python VENV for Linux users). |
|
|
|
|
| Have fun! :D |
|
|
| # Recommendation of newer model - Apex 1.5 Coder |
| Don't get confused by the name - "Coder" doesn't mean, that the model can code very well! But it is finetuned based on this finetuned model (Apex 1.5 Instruct) with CodeAlpaca. |
| Link: https://huggingface.co/LH-Tech-AI/Apex-1.5-Coder-Instruct-350M/ - Have fun with it - it is the newest and best model from LH-Tech AI. |
