Model evaluation¶
Here are the commands you can run to reproduce our evaluation numbers.
The commands below are for OpenMath-2-Llama3.1-8b model as an example.
We assume you have /workspace defined in your cluster config and are
executing all commands from that folder locally. Change all commands accordingly
if running on slurm or using different paths.
Download models¶
Get the model from HF. E.g.
pip install -U "huggingface_hub[cli]"
huggingface-cli download nvidia/OpenMath2-Llama3.1-8B --local-dir OpenMath2-Llama3.1-8B
Convert to TensorRT-LLM¶
Convert the model to TensorRT-LLM format. This is optional, but highly recommended for more exact
results and faster inference. If you skip it, replace --server_type trtllm with --server-type vllm
in the commands below and change model path to /workspace/OpenMath2-Llama3.1-8B. You might also need
to set smaller batch size for vllm.
ns convert \
--cluster=local \
--input_model=/workspace/OpenMath2-Llama3.1-8B \
--output_model=/workspace/openmath2-llama3.1-8b-trtllm \
--convert_from=hf \
--convert_to=trtllm \
--model_type=llama \
--num_gpus=1 \
--hf_model_name=nvidia/OpenMath2-Llama3.1-8B
Change the number of GPUs if you have more than 1 (required for 70B model).
Prepare evaluation data¶
Run greedy decoding¶
ns eval \
--cluster=local \
--model=/workspace/openmath2-llama3.1-8b-trtllm \
--server_type=trtllm \
--output_dir=/workspace/openmath2-llama3.1-8b-eval \
--benchmarks=aime24:0,amc23:0,math:0,gsm8k:0,omni-math:0 \
--server_gpus=1 \
--num_jobs=1 \
++prompt_template=llama3-instruct \
++batch_size=512 \
++inference.tokens_to_generate=4096
If running on slurm, you can set --num_jobs to a bigger number of -1 to run
each benchmark in a separate node. The number of GPUs need to match what you used
in the conversion command.
After the generation is done, we want to run LLM-as-a-judge evaluation to get more
accurate numbers than symbolic comparison. You need to define OPENAI_API_KEY for
the command below to work.
ns llm_math_judge \
--cluster=local \
--model=gpt-4o \
--server_type=openai \
--server_address=https://api.openai.com/v1 \
--input_files="/workspace/openmath2-llama3.1-8b-eval/eval-results/**/output*.jsonl"
Finally, to print the metrics run
This should print the metrics including both symbolic and judge evaluation. The judge is typically more accurate.
------------------------------------------------- aime24 ------------------------------------------------
evaluation_mode | num_entries | symbolic_correct | judge_correct | both_correct | any_correct | no_answer
greedy | 30 | 10.00 | 10.00 | 10.00 | 10.00 | 6.67
------------------------------------------------- gsm8k -------------------------------------------------
evaluation_mode | num_entries | symbolic_correct | judge_correct | both_correct | any_correct | no_answer
greedy | 1319 | 90.75 | 91.70 | 90.75 | 91.70 | 0.00
----------------------------------------------- omni-math -----------------------------------------------
evaluation_mode | num_entries | symbolic_correct | judge_correct | both_correct | any_correct | no_answer
greedy | 4428 | 18.97 | 22.22 | 18.11 | 23.08 | 2.55
-------------------------------------------------- math -------------------------------------------------
evaluation_mode | num_entries | symbolic_correct | judge_correct | both_correct | any_correct | no_answer
greedy | 5000 | 67.70 | 68.10 | 67.50 | 68.30 | 1.36
------------------------------------------------- amc23 -------------------------------------------------
evaluation_mode | num_entries | symbolic_correct | judge_correct | both_correct | any_correct | no_answer
greedy | 40 | 32.50 | 40.00 | 32.50 | 40.00 | 0.00
The numbers may vary by 1-2% depending on the server type, number of GPUs and batch size used.
Run majority voting¶
ns eval \
--cluster=local \
--model=/workspace/openmath2-llama3.1-8b-trtllm \
--server_type=trtllm \
--output_dir=/workspace/openmath2-llama3.1-8b-eval \
--benchmarks=aime24:256,amc23:256,math:256,gsm8k:256,omni-math:256 \
--server_gpus=1 \
--num_jobs=1 \
--skip_greedy \
++prompt_template=llama3-instruct \
++batch_size=512 \
++inference.tokens_to_generate=4096
This will take a very long time unless you run on slurm cluster. After the generation is done, you will be able to see symbolic scores right away. You can evaluate with the judge by first creating new files with majority answers. E.g. for "math" benchmark run
python -m nemo_skills.evaluation.fill_majority_answer \
++input_files="./openmath2-llama3.1-8b-eval/eval-results/math/output-rs*.jsonl" \
++fill_key=predicted_answer
This will replace predicted_answer in all files with majority answer.
After that, let's copy just a single of those files into a new folder so that we can run the llm-judge pipeline on them.
mkdir -p ./openmath2-llama3.1-8b-eval/eval-results-majority/math
cp ./openmath2-llama3.1-8b-eval/eval-results/math/output-rs0.jsonl ./openmath2-llama3.1-8b-eval/eval-results-majority/math/
Repeat the above steps for all benchmarks. Now we are ready to run the judge pipeline and summarize results
after it is finished. You need to define OPENAI_API_KEY for the command below to work.
ns llm_math_judge \
--cluster=local \
--model=gpt-4o \
--server_type=openai \
--server_address=https://api.openai.com/v1 \
--input_files="/workspace/openmath2-llama3.1-8b-eval/eval-results-majority/**/output*.jsonl"
This will print majority results (they will be labeled as majority@1 since we fused them into a single file).
You can also ignore the symbolic score as it's not accurate anymore after we filled majority answers.