|
--- |
|
datasets: |
|
- OpenAssistant/oasst1 |
|
--- |
|
|
|
<p><strong><font size="5">Information</font></strong></p> |
|
OpenAssistant-Llama-13B-4-bit working with GPTQ versions used in Oobabooga's Text Generation Webui and KoboldAI. |
|
<br>This was made using <a href="https://huggingface.co/serpdotai/llama-oasst-lora-13B">Serpdotai's Open Assistant 13b LoRa trained for 4 epochs</a> using Open Assistant's dataset.</p> |
|
|
|
<code>python llama.py /KoboldAI/repos/gptq/llama13b-oasst-4-epochs-lora c4 --wbits 4 --true-sequential --groupsize 128 --save_safetensors llama13b-oasst-4-epochs-lora-4bit-128g.safetensors</code> |
|
|
|
|
|
|
|
<p><strong><font size="5">Benchmarks</font></strong></p> |
|
|
|
<p><strong><font size="4">--true-sequential --groupsize 128</font></strong></p> |
|
|
|
<strong>Wikitext2</strong>: 5.380471229553223 |
|
|
|
<strong>Ptb-New</strong>: 31.921072006225586 |
|
|
|
<strong>C4-New</strong>: 7.140256881713867 |
|
|
|
<strong>Note</strong>: This version uses <i>--groupsize 128</i>, resulting in better evaluations. However, it consumes more VRAM. |
