MetaIX's picture
Update README.md
d9af3f8
<p><strong><font size="5">Information</font></strong></p>
Alpaca 30B 4-bit working with GPTQ versions used in Oobabooga's Text Generation Webui and KoboldAI.
<p>Quantized using <i>--true-sequential</i> and <i>--groupsize 128</i> optimizations.</p>
This was made using Chansung's 30B Alpaca Lora: https://huggingface.co/chansung/alpaca-lora-30b
<p><strong><font size="5">Update 04.06.2023</font></strong></p>
<p>This is a more recent merge of Chansung's Alpaca Lora which was updated using the clean alpaca dataset as of 04/06/2023 with refined training parameters</p>
<p><strong>Training Parameters</strong></p>
<ul><li>num_epochs=10</li><li>cutoff_len=512</li><li>group_by_length</li><li>lora_target_modules='[q_proj,k_proj,v_proj,o_proj]'</li><li>lora_r=16</li><li>micro_batch_size=8</li></ul>
<p><strong><font size="5">Benchmarks</font></strong></p>
<strong>Wikitext2</strong>: 4.473957061767578
<strong>Ptb-New</strong>: 8.682597160339355
<strong>C4-New</strong>: 6.517213344573975
<strong>Note</strong>: This version uses <i>--groupsize 128</i>, resulting in better evaluations. However, it consumes more VRAM.