6e10d98a8d
* add llm-ppl workflow * update the DATASET_DIR * test multiple precisions * modify nightly test * match the updated ppl code * add matrix.include * fix the include error * update the include * add more model * update the precision of include * update nightly time and add more models * fix the workflow_dispatch description, change default model of pr and modify the env * modify workflow_dispatch language options * modify options * modify language options * modeify workflow_dispatch type * modify type * modify the type of language * change seq_len type * fix some typos * revert changes to stress_test.txt |
||
|---|---|---|
| .. | ||
| bigdl_llm.py | ||
| fp16.csv | ||
| harness_csv_to_html.py | ||
| harness_to_leaderboard.py | ||
| make_table_and_csv.py | ||
| README.md | ||
| run_llb.py | ||
| run_multi_llb.py | ||
Harness Evaluation
Harness evaluation allows users to eaisly get accuracy on various datasets. Here we have enabled harness evaluation with BigDL-LLM under Open LLM Leaderboard settings. Before running, make sure to have bigdl-llm installed.
Install Harness
git clone https://github.com/EleutherAI/lm-evaluation-harness.git
cd lm-evaluation-harness
git checkout e81d3cc
pip install -e .
Run
run python run_llb.py. run_llb.py combines some arguments in main.py to make evaluations easier. The mapping of arguments is defined as a dict in llb.py.
Evaluation on CPU
python run_llb.py --model bigdl-llm --pretrained /path/to/model --precision nf3 sym_int4 nf4 --device cpu --tasks hellaswag arc mmlu truthfulqa --batch 1 --no_cache
Evaluation on Intel GPU
python run_llb.py --model bigdl-llm --pretrained /path/to/model --precision nf3 sym_int4 nf4 --device xpu --tasks hellaswag arc mmlu truthfulqa --batch 1 --no_cache
Evaluation using multiple Intel GPU
python run_multi_llb.py --model bigdl-llm --pretrained /path/to/model --precision nf3 sym_int4 nf4 --device xpu:0,2,3 --tasks hellaswag arc mmlu truthfulqa --batch 1 --no_cache
Taking example above, the script will fork 3 processes, each for one xpu, to execute the tasks.
Results
We follow Open LLM Leaderboard to record our metrics, acc_norm for hellaswag and arc_challenge, mc2 for truthful_qa and acc for mmlu. For mmlu, there are 57 subtasks which means users may need to average them manually to get final result.