Skip to content

Latest commit

 

History

History
45 lines (33 loc) · 1.57 KB

README.md

File metadata and controls

45 lines (33 loc) · 1.57 KB

SmolLM evaluation scripts

We're using the LightEval library to benchmark our models.

Check out the quick tour to configure it to your own hardware and tasks.

Setup

Use conda/venv with python>=3.10.

Adjust the pytorch installation according to your environment:

pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu121

For reproducibility, we recommend fixed versions of the libraries:

pip install -r requirements.txt

Running the evaluations

SmolLM2 base models

lighteval accelerate \
  --model_args "pretrained=HuggingFaceTB/SmolLM2-1.7B,revision=main,dtype=bfloat16,vllm,gpu_memory_utilisation=0.8,max_model_length=2048" \
  --custom_tasks "tasks.py" --tasks "smollm2_base.txt" --output_dir "./evals" --save_details

SmolLM2 instruction-tuned models

(note the --use_chat_template flag)

lighteval accelerate \
  --model_args "pretrained=HuggingFaceTB/SmolLM2-1.7B-Instruct,revision=main,dtype=bfloat16,vllm,gpu_memory_utilisation=0.8,max_model_length=2048" \
  --custom_tasks "tasks.py" --tasks "smollm2_instruct.txt" --use_chat_template --output_dir "./evals" --save_details

MATH and other extra tasks

lighteval accelerate \
  --model_args "pretrained=HuggingFaceTB/SmolLM2-1.7B-Instruct,revision=main,dtype=bfloat16,vllm,gpu_memory_utilisation=0.8,max_model_length=4096" \
  --custom_tasks "tasks.py" --tasks "custom|math|4|1" --use_chat_template --output_dir "./evals" --save_details