catid's picture
results
05b6b0c

AI Model Name: Llama 3 8B "Built with Meta Llama 3" https://llama.meta.com/llama3/license/

Full walkthrough to reproduce these results here: https://github.com/catid/AQLM/blob/main/catid_readme.md

Baseline evaluation results:

hf (pretrained=meta-llama/Meta-Llama-3-8B-Instruct), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 16
|    Tasks    |Version|Filter|n-shot| Metric |Value |   |Stderr|
|-------------|------:|------|-----:|--------|-----:|---|-----:|
|winogrande   |      1|none  |     0|acc     |0.7198|±  |0.0126|
|piqa         |      1|none  |     0|acc     |0.7873|±  |0.0095|
|             |       |none  |     0|acc_norm|0.7867|±  |0.0096|
|hellaswag    |      1|none  |     0|acc     |0.5767|±  |0.0049|
|             |       |none  |     0|acc_norm|0.7585|±  |0.0043|
|arc_easy     |      1|none  |     0|acc     |0.8140|±  |0.0080|
|             |       |none  |     0|acc_norm|0.7971|±  |0.0083|
|arc_challenge|      1|none  |     0|acc     |0.5290|±  |0.0146|
|             |       |none  |     0|acc_norm|0.5674|±  |0.0145|

This repo evaluation results (AQLM with global fine-tuning):

hf (pretrained=catid/cat-llama-3-8b-instruct-aqlm), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 16
|    Tasks    |Version|Filter|n-shot| Metric |Value |   |Stderr|
|-------------|------:|------|-----:|--------|-----:|---|-----:|
|winogrande   |      1|none  |     0|acc     |0.7119|±  |0.0127|
|piqa         |      1|none  |     0|acc     |0.7807|±  |0.0097|
|             |       |none  |     0|acc_norm|0.7824|±  |0.0096|
|hellaswag    |      1|none  |     0|acc     |0.5716|±  |0.0049|
|             |       |none  |     0|acc_norm|0.7539|±  |0.0043|
|arc_easy     |      1|none  |     0|acc     |0.8152|±  |0.0080|
|             |       |none  |     0|acc_norm|0.7866|±  |0.0084|
|arc_challenge|      1|none  |     0|acc     |0.5043|±  |0.0146|
|             |       |none  |     0|acc_norm|0.5555|±  |0.0145|

To reproduce evaluation results:

git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness

conda create -n lmeval python=3.10 -y && conda activate lmeval
pip install -e .
pip install accelerate aqlm"[gpu,cpu]"

accelerate launch lm_eval --model hf \
    --model_args pretrained=catid/cat-llama-3-8b-instruct-aqlm \
    --tasks winogrande,piqa,hellaswag,arc_easy,arc_challenge \
    --batch_size 16

You can run this model as a transformers model using https://github.com/oobabooga/text-generation-webui