QoQ-g128 Llama3-70B-Instruct Results

[ethxnp]

I ran python -m lmquant.llm.run projects/llm/configs/llm.yaml projects/llm/configs/qoq/g128.yaml --mod│ el-name llama3-instruct-70b --model-path meta-llama/Meta-Llama-3-70B-Instruct --smooth-xw-alpha 0.3 --smooth-xw-beta 0.7 --smooth-yx-str│ ategy GridSearch --smooth-yx-beta " -2" --save-model

These were the results:

24-06-03 19:43:57 | I |         |  Task  |Version|    Metric     |Value |   |Stderr|                                                    │
24-06-03 19:43:57 | I |         |--------|------:|---------------|-----:|---|-----:|                                                    │
24-06-03 19:43:57 | I |         |wikitext|      1|word_perplexity|5.5105|±  |5.5105|

The perplexity seems rather high, do you have recommendations for different parameters to try? I will upload to HF later and share the link in case it's helpful.