Skip to content

Latest commit

 

History

History
137 lines (103 loc) · 5.49 KB

evaluation.md

File metadata and controls

137 lines (103 loc) · 5.49 KB

Model evaluation

Quick start

Make sure to complete prerequisites before proceeding. Please note that ~1% difference in accuracy is expected when running inference on different GPU types or with different inference frameworks.

  1. Download one of our models or get some other checkpoint.

  2. Convert the model to the right format if required.

  3. Run the evaluation (assuming one of our finetuned models, nemo inference, gsm8k greedy decoding)

    python pipeline/run_eval.py \
  --model_path <path to .nemo> \
  --server_type nemo \
  --output_dir ./test-results \
  --benchmarks gsm8k:0 \
  --num_gpus <number of GPUs on your machine/cluster node> \
  --num_jobs 1 \
  +prompt=openmathinstruct/sft \
  ++prompt.few_shot_examples.num_few_shots=0 \
  ++split_name=test

    If you want to evaluate a model that was not finetuned through our pipeline, but still allow it to use Python interpreter, you can show it a couple of few-shot examples

    +prompt=openmathinstruct/base \
++prompt.few_shot_examples.examples_type=gsm8k_text_with_code \
++prompt.few_shot_examples.num_few_shots=5

    If you need to, change the batch size with batch_size=<X> argument.

  4. Compute metrics

    python pipeline/compute_metrics.py \
  --prediction_jsonl_files ./test-results/gsm8k/output-greedy.jsonl \
  --benchmark gsm8k

    If you evaluated multiple benchmarks or used multiple samples per benchmark, you can also run the following script to summarize all available metrics.

    python pipeline/summarize_results.py ./test-results

Read on to learn details about how evaluation works!

Details

Let's break down what pipeline/run_eval.py is doing.

  • Starts a local sandbox which will handle code execution requests.
  • Starts an LLM server in a docker container (defined in the NEMO_SKILLS_CONFIG file).
  • Waits for the sandbox and server to start.
  • Runs nemo_skills/inference/generate_solutions.py to generate solutions for all benchmarks requested (potentially running multiple samples per benchmark).
  • Runs nemo_skills/evaluation/evaluate_results.py on each of the generated output files.
  • If running in a Slurm cluster, you can parallelize evaluation across multiple nodes. You can also customize any of the parameters of evaluation - all extra arguments of the run_eval.py will be passed directly to the generate_solutions.py script.

Here is an example of how to manually reproduce the call to run_eval.py script from the quick start section.

  1. Start a sandbox. This will block your shell, so either run in the background or make sure you can open another shell on the same machine:

    ./nemo_skills/code_execution/local_sandbox/start_local_sandbox.sh

    Get the IP of the sandbox by running

    docker inspect -f '{{range.NetworkSettings.Networks}}{{.IPAddress}}{{end}}' `docker ps -a | grep local-sandbox  | awk '{print $1}'`

  2. Start an LLM server. The commands differ based on the server type. Here is an example for starting NeMo-based inference server. Make sure to run this from the root of the repository. Same as above, this will block your shell.

    docker run --rm --gpus all --ipc=host -v `pwd`:/code -v <path to the .nemo model>:/model igitman/nemo-skills-sft:0.3.0 \
bash -c 'PYTHONPATH=/code python /code/nemo_skills/inference/server/serve_nemo.py \
  gpt_model_file=/model \
  trainer.devices=<number of GPUs> \
  tensor_model_parallel_size=<number of GPUs> \
  ++sandbox.host=<Sandbox IP from the step above>'

    Wait until you see "Running on " message and make a note of this IP.

    If you want to use TensorRT-LLM server instead, you can run the following command

    docker run --rm --gpus all --ipc=host -v `pwd`:/code -v <path to the trtllm model>:/model igitman/nemo-skills-trtllm:0.3.2 \
bash -c 'export PYTHONPATH=/code && \
mpirun -n <number of GPUs> --allow-run-as-root python /code/nemo_skills/inference/server/serve_trt.py --model_path=/model'

  3. Run the generation command. Customize as necessary (running with --help will show the details)

    python nemo_skills/inference/generate_solutions.py \
  output_file=./test-results/gsm8k/output-greedy.jsonl \
  +prompt=openmathinstruct/sft \
  ++dataset=gsm8k \
  ++split_name=test \
  ++server.server_type=nemo \
  ++server.host=<IP from the step above> \
  ++sandbox.host=<Sandbox IP from the sandbox launch step>

  4. Run the evaluation command. Note that you need to provide a sandbox IP, because evaluation is running in the sandbox.

    python nemo_skills/evaluation/evaluate_results.py \
  prediction_jsonl_files=./test-results/gsm8k/output-greedy.jsonl \
  ++sandbox.host=<Sandbox IP>

After this you would typically follow up with the same command to compute metrics as in the quick start.

Typical customizations

To customize the prompt template for the model, create a new .yaml file inside nemo_skills/inference/prompt folder. Have a look at the existing templates there for an example.

You can run python nemo_skills/inference/generate_solutions.py --help to see other available customization options.