Add GPQA Diamond and fix evaluation deps

README.md

@@ -56,17 +56,17 @@ uv venv openr1 --python 3.11 && source openr1/bin/activate && uv pip install --u
 Next, install vLLM:
 
 ```shell
-uv pip install vllm>=0.7.0
+uv pip install vllm==0.7.1
 
 # For CUDA 12.1
-pip install vllm>=0.7.0 --extra-index-url https://download.pytorch.org/whl/cu121
+uv pip install vllm==0.7.1 --extra-index-url https://download.pytorch.org/whl/cu121 --index-strategy unsafe-best-match
 export LD_LIBRARY_PATH=$(python -c "import site; print(site.getsitepackages()[0] + '/nvidia/nvjitlink/lib')"):$LD_LIBRARY_PATH
 ```
 
 This will also install PyTorch `v2.5.1` and it is **very important** to use this version since the vLLM binaries are compiled for it. You can then install the remaining dependencies for your specific use case via `pip install -e .[LIST OF MODES]`. For most contributors, we recommend:
 
 ```shell
-pip install -e ".[dev]"
+GIT_LFS_SKIP_SMUDGE=1 uv pip install -e ".[dev]" --link-mode=copy
 ```
 
 Next, log into your Hugging Face and Weights and Biases accounts as follows:
@@ -142,16 +142,33 @@ We use `lighteval` to evaluate models, with custom tasks defined in `src/open_r1
 ```shell
 MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
 MODEL_ARGS="pretrained=$MODEL,dtype=float16,max_model_length=32768,gpu_memory_utilisation=0.8"
-TASK=aime24
 OUTPUT_DIR=data/evals/$MODEL
 
+# AIME 2024
+TASK=aime24
+lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
+    --custom-tasks src/open_r1/evaluate.py \
+    --use-chat-template \
+    --output-dir $OUTPUT_DIR
+
+# MATH-500
+TASK=math_500
+lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
+    --custom-tasks src/open_r1/evaluate.py \
+    --use-chat-template \
+    --output-dir $OUTPUT_DIR
+
+# GPQA Diamond
+TASK=gpqa:diamond
 lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
     --custom-tasks src/open_r1/evaluate.py \
     --use-chat-template \
-    --system-prompt="Please reason step by step, and put your final answer within \boxed{}." \
     --output-dir $OUTPUT_DIR 
 ```
 
+> [!IMPORTANT]
+> You must set `max_model_length=32768` in the `vllm` command to align with the `generation_size` we define per eval. Without this, `lighteval` will throw an error.
+
 To increase throughput across multiple GPUs, use _data parallel_ as follows:
 
 ```shell
@@ -164,7 +181,6 @@ OUTPUT_DIR=data/evals/$MODEL
 lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
     --custom-tasks src/open_r1/evaluate.py \
     --use-chat-template \
-    --system-prompt="Please reason step by step, and put your final answer within \boxed{}." \
     --output-dir $OUTPUT_DIR 
 ```
 
@@ -181,50 +197,94 @@ export VLLM_WORKER_MULTIPROC_METHOD=spawn
 lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
     --custom-tasks src/open_r1/evaluate.py \
     --use-chat-template \
-    --system-prompt="Please reason step by step, and put your final answer within \boxed{}." \
     --output-dir $OUTPUT_DIR 
 ```
 
 You can also launch an evaluation with `make evaluate`, specifying the model, task, and optionally the parallelism technique and number of GPUs.
 
 To evaluate on a single GPU:
+
 ```shell
 make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24
 ```
 
 To use Data Parallelism:
+
 ```shell
 make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24 PARALLEL=data NUM_GPUS=8
 ```
 
 To use Tensor Parallelism:
+
 ```shell
 make evaluate MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-32B TASK=aime24 PARALLEL=tensor NUM_GPUS=8
 ```
-## Reproducing Deepseek's evaluation results on MATH-500
-We are able to reproduce Deepseek's reported results on the MATH-500 Benchmark:
-| Model                      | MATH-500 (HF lighteval) | MATH-500 (DeepSeek Reported) |
-| :-------------------------- | :-------: | :----------------------------: |
-| DeepSeek-R1-Distill-Qwen-1.5B  |  81.6   |              83.9              |
-| DeepSeek-R1-Distill-Qwen-7B    |  91.8   |              92.8              |
-| DeepSeek-R1-Distill-Qwen-14B   |  94.2   |              93.9              |
-| DeepSeek-R1-Distill-Qwen-32B   |  95.0   |              94.3              |
-| DeepSeek-R1-Distill-Llama-8B   |  85.8   |              89.1              |
-| DeepSeek-R1-Distill-Llama-70B  |  93.4   |              94.5              |
 
+## Reproducing Deepseek's evaluation results
+
+### MATH-500
 
+We are able to reproduce Deepseek's reported results on the MATH-500 benchmark within ~1 standard deviation:
+
+| Model                         | MATH-500 (🤗 LightEval) | MATH-500 (DeepSeek Reported) |
+|:------------------------------|:-----------------------:|:----------------------------:|
+| DeepSeek-R1-Distill-Qwen-1.5B |          81.8           |             83.9             |
+| DeepSeek-R1-Distill-Qwen-7B   |          91.8           |             92.8             |
+| DeepSeek-R1-Distill-Qwen-14B  |          94.2           |             93.9             |
+| DeepSeek-R1-Distill-Qwen-32B  |          95.0           |             94.3             |
+| DeepSeek-R1-Distill-Llama-8B  |          85.8           |             89.1             |
+| DeepSeek-R1-Distill-Llama-70B |          93.4           |             94.5             |
 
 To reproduce these results use the following command:
+
+```shell
+NUM_GPUS=8
+MODEL=deepseek-ai/{model_name}
+MODEL_ARGS="pretrained=$MODEL,dtype=float16,max_model_length=32768,gpu_memory_utilisation=0.8,tensor_parallel_size=$NUM_GPUS"
+OUTPUT_DIR=data/evals/$MODEL
+
+lighteval vllm $MODEL_ARGS "custom|math_500|0|0" \
+    --custom-tasks src/open_r1/evaluate.py \
+    --use-chat-template \
+    --output-dir $OUTPUT_DIR
+```
+
+Alternatively, you can launch Slurm jobs as follows:
+
 ```shell
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B math_500
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Qwen-7B math_500
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Qwen-14B math_500
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Qwen-32B math_500 tp
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Llama-8B math_500
-sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Llama-70B math_500 tp
+python scripts/run_benchmarks.py --model-id={model_id}  --benchmarks math_500
 ```
 
+### GPQA Diamond
+
+We are able to reproduce Deepseek's reported results on the GPQA Diamond benchmark within ~1 standard deviation:
+
+| Model                         | GPQA Diamond (🤗 LightEval) | GPQA Diamond (DeepSeek Reported) |
+|:------------------------------|:---------------------------:|:--------------------------------:|
+| DeepSeek-R1-Distill-Qwen-1.5B |            33.33            |               33.8               |
+| DeepSeek-R1-Distill-Qwen-7B   |            48.48            |               49.1               |
+| DeepSeek-R1-Distill-Qwen-14B  |            55.56            |               59.1               |
+| DeepSeek-R1-Distill-Qwen-32B  |            58.59            |               62.1               |
+| DeepSeek-R1-Distill-Llama-8B  |            51.01            |               49.0               |
+| DeepSeek-R1-Distill-Llama-70B |            65.15            |               65.2               |
+
+To reproduce these results use the following command:
+
+```shell
+NUM_GPUS=8
+MODEL=deepseek-ai/{model_name}
+MODEL_ARGS="pretrained=$MODEL,dtype=float16,max_model_length=32768,gpu_memory_utilisation=0.8,tensor_parallel_size=$NUM_GPUS"
+OUTPUT_DIR=data/evals/$MODEL
+
+lighteval vllm $MODEL_ARGS "custom|gpqa:diamond|0|0" \
+    --custom-tasks src/open_r1/evaluate.py \
+    --use-chat-template \
+    --output-dir $OUTPUT_DIR
+```
 
+```shell
+python scripts/run_benchmarks.py --model-id={model_id}  --benchmarks gpqa
+```
 
 ## Data generation
 

setup.py

@@ -53,17 +53,17 @@
     "huggingface-hub[cli]>=0.19.2,<1.0",
     "isort>=5.12.0",
     "liger_kernel==0.5.2",
-    "lighteval @ git+https://github.com/huggingface/lighteval.git@0e462692436e1f0575bdb4c6ef63453ad9bde7d4#egg=lighteval[math]",
-    "math-verify>=0.3.3",  # Used for math verification in grpo
+    "lighteval @ git+https://github.com/huggingface/lighteval.git@3c9b0c9dde6718b23ef5b0f4960355f0d494bdfc#egg=lighteval[math]",
+    "math-verify==0.5.2",  # Used for math verification in grpo
     "packaging>=23.0",
     "parameterized>=0.9.0",
     "pytest",
     "safetensors>=0.3.3",
     "sentencepiece>=0.1.99",
-    "torch>=2.5.1",
+    "torch==2.5.1",
     "transformers @ git+https://github.com/huggingface/transformers.git@main",
     "trl @ git+https://github.com/huggingface/trl.git@main",
-    "vllm>=0.7.1",
+    "vllm==0.7.1",
     "wandb>=0.19.1",
 ]
 

slurm/evaluate.slurm

@@ -1,55 +1,75 @@
 #!/bin/bash
-#SBATCH --job-name=open-r1-evaluate
-#SBATCH --nodes=1
 #SBATCH --ntasks-per-node=1
-#SBATCH --exclusive
 #SBATCH --gres=gpu:8
-#SBATCH --partition=hopper-prod 
-#SBATCH --time=01:59:00
-#SBATCH --output=./logs/evaluate/%x-%j.out
-#SBATCH --err=./logs/evaluate/%x-%j.err
-
-# Usage: sbatch slurm/evaluate.slurm deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B aime24
+#SBATCH --partition=hopper-prod
+#SBATCH --output=./logs/%x-%j.out
+#SBATCH --err=./logs/%x-%j.err
+#SBATCH --requeue
 
 set -x -e
-
 source ~/.bashrc
 source openr1/bin/activate
-module load cuda/12.1
-echo "START TIME: $(date)"
-echo "PYTHON ENV: $(which python)"
-
+TASK_NAME=$1
+TASKS=$2
+MODEL_ID=$3
+MODEL_REVISION=$4
+# Optional args
+[ -z "$5"] && TENSOR_PARALLEL=False || TENSOR_PARALLEL=$5
+[ -z "$6"] && TRUST_REMOTE_CODE=False || TRUST_REMOTE_CODE=$6
+# $7 is reserved for system_prompt, see line 51
+NUM_GPUS=$(nvidia-smi -L | wc -l)
 
-NUM_GPUS=8
-MODEL=$1
-TASK=$2
-# Check if a third argument is passed, if it is tp then eval with tensor parallelism. Required for larger models
-if [ -n "$3" ] && [ "$3" == "tp" ]; then
-  MODEL_ARGS="pretrained=$MODEL,dtype=float16,tensor_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
+# Set Whether to use tensor parallelism or data parallelism
+if [ "$TENSOR_PARALLEL" = "True" ]; then
+    # use TP to shard model across NUM_GPUS
+    export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    MODEL_ARGS="pretrained=$MODEL_ID,revision=$MODEL_REVISION,trust_remote_code=$TRUST_REMOTE_CODE,dtype=float16,tensor_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
 else
-  MODEL_ARGS="pretrained=$MODEL,dtype=float16,data_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
+    MODEL_ARGS="pretrained=$MODEL_ID,revision=$MODEL_REVISION,trust_remote_code=$TRUST_REMOTE_CODE,dtype=float16,data_parallel_size=$NUM_GPUS,max_model_length=32768,gpu_memory_utilisation=0.8"
 fi
-OUTPUT_DIR=data/evals/$MODEL
-
-
-# force crashing on nccl issues like hanging broadcast
-export NCCL_ASYNC_ERROR_HANDLING=1
-# export NCCL_DEBUG=INFO
-# export NCCL_DEBUG_SUBSYS=COLL
-# export NCCL_SOCKET_NTHREADS=1
-# export NCCL_NSOCKS_PERTHREAD=1
-# export CUDA_LAUNCH_BLOCKING=1
 
-# Specific configuration optimized for the Hugging Face Compute Cluster
-# Be ye warned this may not work on other clusters!
-module load cuda/12.1
+LM_EVAL_REPO_ID="open-r1/open-r1-eval-leaderboard"
+MODEL_NAME=$(echo $MODEL_ID | sed 's/\//_/g') # replaces / with _
+DETAILS_REPO_ID="open-r1/details-$MODEL_NAME"
+OUTPUT_DIR="eval_results/$MODEL_ID/$MODEL_REVISION/$TASK_NAME"
+# We need this flag since we run this script from training jobs that use DeepSpeed and the env vars get progated which causes errors during evaluation
+ACCELERATE_USE_DEEPSPEED=false
+# Enable fast downloads
+HF_HUB_ENABLE_HF_TRANSFER=1
 
-lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
-    --custom-tasks src/open_r1/evaluate.py \
+echo "Running lighteval script ..."
+echo "Eval results will be saved to $OUTPUT_DIR"
+# Check if "custom" is a substring of TASKS
+if [[ $TASKS == *"custom"* ]]; then
+    echo "Custom task detected. Running custom task evaluation script ..."
+    lighteval vllm $MODEL_ARGS $TASKS \
+    --custom-tasks "src/open_r1/evaluate.py" \
     --use-chat-template \
-    --system-prompt="Please reason step by step, and put your final answer within \boxed{}." \
+    --output-dir $OUTPUT_DIR \
     --save-details \
-    --output-dir $OUTPUT_DIR 
+    ${7:+--system-prompt "$7"}
+else
+    lighteval vllm $MODEL_ARGS $TASKS \
+    --use-chat-template \
+    --output-dir $OUTPUT_DIR \
+    --save-details \
+    ${7:+--system-prompt "$7"}
+fi
+
+OUTPUT_FILEPATHS=$(find $OUTPUT_DIR/results/ -type f \( -name "*.json" \))
+for filepath in $OUTPUT_FILEPATHS; do
+    echo "Uploading $filepath to Hugging Face Hub..."
+    filename=$(basename -- "$filepath")
+    huggingface-cli upload --repo-type space --private $LM_EVAL_REPO_ID $filepath $OUTPUT_DIR/$filename
+done
 
+echo "Uploading details to Hugging Face Hub..."
+DETAILS_FILEPATHS=$(find $OUTPUT_DIR/details/ -type f \( -name "*.parquet" \))
+echo "DETAILS_FILEPATHS: $DETAILS_FILEPATHS"
+TIMESTAMP=$(date +"%Y-%m-%dT%H-%M-%S")
+python src/open_r1/utils/upload_details.py --data_files $DETAILS_FILEPATHS --hub_repo_id $DETAILS_REPO_ID --config_name $MODEL_REVISION.$TASK_NAME.$TIMESTAMP
+
+echo "Cleaning up ..."
+rm -rf $OUTPUT_DIR
 
-echo "END TIME: $(date)"
+echo "Done!"