mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Shuhe WangandGitHub
29b23d057c
* Add the parameter "tokenizer_mode" when using vllm to load the model for AlpacaEval * Update alpaca_farm.sh
26 lines
914 B
Bash
Executable File
26 lines
914 B
Bash
Executable File
# Please make sure OPENAI_API_KEY is set in your environment variables
|
|
|
|
# Use V1 of alpaca farm evaluation.
|
|
export IS_ALPACA_EVAL_2=False
|
|
|
|
# use vllm for generation
|
|
python -m eval.alpaca_farm.run_eval \
|
|
--model_name_or_path ../checkpoints/tulu_v1_7B/ \
|
|
--tokenizer_name_or_path ../checkpoints/tulu_v1_7B/ \
|
|
--save_dir results/alpaca_farm/tulu_v1_7B/ \
|
|
--eval_batch_size 20 \
|
|
--use_vllm \
|
|
--use_chat_format \
|
|
--chat_formatting_function eval.templates.create_prompt_with_tulu_chat_format
|
|
|
|
|
|
# use normal huggingface generation function
|
|
python -m eval.alpaca_farm.run_eval \
|
|
--model_name_or_path ../checkpoints/tulu_v1_7B/ \
|
|
--tokenizer_name_or_path ../checkpoints/tulu_v1_7B/ \
|
|
--save_dir results/alpaca_farm/tulu_v1_7B/ \
|
|
--eval_batch_size 20 \
|
|
--use_chat_format \
|
|
--chat_formatting_function eval.templates.create_prompt_with_tulu_chat_format \
|
|
--load_in_8bit
|