tangger/Search-R1
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add paper scripts

Browse Source
This commit is contained in:
PeterGriffinJin
2025-03-13 13:57:47 +00:00
parent 0ecaf6da76
commit 584ce9deb5
5 changed files with 270 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
README.md
View File

@@ -4,11 +4,10 @@
Through RL (rule-based outcome reward), the 3B **base** LLM (both Qwen2.5-3b-base and Llama3.2-3b-base) develops reasoning and search engine calling abilities all on its own.
Twitter thread: [link](https://x.com/BowenJin13/status/1895544294473109889); Full experiment log: [link](https://wandb.ai/peterjin/Search-R1-open)
Paper: [link](https://arxiv.org/pdf/2503.09516); Model and data: [link](https://huggingface.co/collections/PeterJinGo/search-r1-67d1a021202731cb065740f5); Twitter thread: [link](https://x.com/BowenJin13/status/1895544294473109889); Full experiment log 1: [link](https://wandb.ai/peterjin/Search-R1-open); Full experiment log 2: [link](hhttps://wandb.ai/peterjin/Search-R1-nq_hotpotqa_train/)
Paper: [link](); Model and data: [link](https://huggingface.co/collections/PeterJinGo/search-r1-67d1a021202731cb065740f5);
You can refer to this [link](https://github.com/PeterGriffinJin/Search-R1/tree/main/scripts/nq_hotpotqa) for detailed instructions on reproducing the results from the paper.
The paper will be released soon!
![single-turn](public/single-turn.png)
@@ -166,11 +165,10 @@ You can refer to ```search_r1/search/retriever_server.py``` for an example of la
The concept of Search-R1 is inspired by [Deepseek-R1](https://github.com/deepseek-ai/DeepSeek-R1) and [TinyZero](https://github.com/Jiayi-Pan/TinyZero/tree/main).
Its implementation is built upon [veRL](https://github.com/volcengine/verl) and [RAGEN](https://github.com/ZihanWang314/RAGEN/tree/main).
We sincerely appreciate the efforts of these teams for their contributions to open-source research and development.
We thank Jinsung Yoon and Sercan Arik for insightful discussions.
## Citations
To be added
```bibtex
@misc{jin2025searchr1,
title = {Search-R1: Train your LLMs to reason and call a search engine with reinforcement learning},

26
scripts/nq_hotpotqa/README.md Normal file
View File

@@ -0,0 +1,26 @@
## Reproduce the paper results
### Download the dataset
```bash
huggingface-cli download --repo-type dataset PeterJinGo/nq_hotpotqa_train --local-dir $WORK_DIR/data/hotpot_qa
```
### Run PPO training
```bash
bash train_ppo.sh
```
### Run GRPO training
```bash
bash train_ppo.sh
```
### Run evaluation
```bash
bash evaluate.sh
```
You can change ```$BASE_MODEL``` to the path of the model you would loike to evaluate.

65
scripts/nq_hotpotqa/evaluate.sh Normal file
View File

@@ -0,0 +1,65 @@
data_name=nq_hotpotqa_train
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export DATA_DIR=data/${data_name} # first download the data from https://huggingface.co/datasets/PeterJinGo/nq_hotpotqa_train
export BASE_MODEL=""
# set -x
export VLLM_ATTENTION_BACKEND=XFORMERS # vllm + qwen2-7b with flash_attn has some issues
# max_prompt_length = (config['training']['max_start_length'] + config['training']['max_response_length'] * (config['training']['max_turns'] - 1) + config['training']['max_obs_length'] * config['training']['max_turns'])
PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
data.train_files=$DATA_DIR/train.parquet \
data.val_files=$DATA_DIR/test.parquet \
data.train_data_num=null \
data.val_data_num=null \
data.train_batch_size=512 \
data.val_batch_size=256 \
data.max_prompt_length=4096 \
data.max_response_length=500 \
data.max_start_length=2048 \
data.max_obs_length=500 \
data.shuffle_train_dataloader=True \
algorithm.adv_estimator=gae \
actor_rollout_ref.model.path=$BASE_MODEL \
actor_rollout_ref.actor.optim.lr=1e-6 \
actor_rollout_ref.model.enable_gradient_checkpointing=true \
actor_rollout_ref.model.use_remove_padding=True \
actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=0.95 \
actor_rollout_ref.actor.ppo_mini_batch_size=256 \
actor_rollout_ref.actor.ppo_micro_batch_size=64 \
actor_rollout_ref.actor.fsdp_config.param_offload=true \
actor_rollout_ref.actor.fsdp_config.grad_offload=true \
actor_rollout_ref.actor.fsdp_config.optimizer_offload=true \
actor_rollout_ref.rollout.log_prob_micro_batch_size=128 \
actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
actor_rollout_ref.rollout.name=vllm \
actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
actor_rollout_ref.ref.log_prob_micro_batch_size=128 \
actor_rollout_ref.ref.fsdp_config.param_offload=True \
actor_rollout_ref.rollout.n_agent=1 \
actor_rollout_ref.rollout.temperature=1 \
actor_rollout_ref.actor.state_masking=true \
critic.optim.lr=1e-5 \
critic.model.use_remove_padding=True \
critic.optim.lr_warmup_steps_ratio=0.05 \
critic.model.path=$BASE_MODEL \
critic.model.enable_gradient_checkpointing=true \
critic.ppo_micro_batch_size=8 \
critic.model.fsdp_config.param_offload=true \
critic.model.fsdp_config.grad_offload=true \
critic.model.fsdp_config.optimizer_offload=true \
algorithm.kl_ctrl.kl_coef=0.001 \
algorithm.no_think_rl=false \
trainer.critic_warmup=0 \
trainer.logger=[] \
+trainer.val_only=true \
+trainer.val_before_train=true \
trainer.default_hdfs_dir=null \
trainer.n_gpus_per_node=8 \
trainer.nnodes=1 \
max_turns=4 \
retriever.url="http://127.0.0.1:8000/retrieve" \
retriever.topk=3

84
scripts/nq_hotpotqa/train_grpo.sh Normal file
View File

@@ -0,0 +1,84 @@
data_name=nq_hotpotqa_train
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export DATA_DIR=data/${data_name} # first download the data from https://huggingface.co/datasets/PeterJinGo/nq_hotpotqa_train
WAND_PROJECT="Search-R1"
export BASE_MODEL='meta-llama/Llama-3.2-3B'
export EXPERIMENT_NAME=${data_name}-search-r1-grpo-llama3.2-3b-em
# export BASE_MODEL='meta-llama/Llama-3.2-3B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-llama3.2-3b-it-em
# export BASE_MODEL='meta-llama/Llama-3.1-8B'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-llama3.1-8b-em
# export BASE_MODEL='meta-llama/Llama-3.1-8B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-llama3.1-8b-it-em
# export BASE_MODEL='Qwen/Qwen2.5-3B'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-qwen2.5-3b-em
# export BASE_MODEL='Qwen/Qwen2.5-3B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-qwen2.5-3b-it-em
# export BASE_MODEL='Qwen/Qwen2.5-7B'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-qwen2.5-7b-em
# export BASE_MODEL='Qwen/Qwen2.5-7B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-grpo-qwen2.5-7b-it-em
# set -x
export VLLM_ATTENTION_BACKEND=XFORMERS # vllm + qwen2-7b with flash_attn has some issues
# max_prompt_length = (config['training']['max_start_length'] + config['training']['max_response_length'] * (config['training']['max_turns'] - 1) + config['training']['max_obs_length'] * config['training']['max_turns'])
PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
data.train_files=$DATA_DIR/train.parquet \
data.val_files=$DATA_DIR/test.parquet \
data.train_data_num=null \
data.val_data_num=null \
data.train_batch_size=512 \
data.val_batch_size=256 \
data.max_prompt_length=4096 \
data.max_response_length=500 \
data.max_start_length=2048 \
data.max_obs_length=500 \
data.shuffle_train_dataloader=True \
algorithm.adv_estimator=grpo \
actor_rollout_ref.model.path=$BASE_MODEL \
actor_rollout_ref.model.enable_gradient_checkpointing=true \
actor_rollout_ref.model.use_remove_padding=True \
actor_rollout_ref.actor.optim.lr=1e-6 \
actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=0.95 \
actor_rollout_ref.actor.use_kl_loss=true \
actor_rollout_ref.actor.ppo_mini_batch_size=256 \
actor_rollout_ref.actor.ppo_micro_batch_size=64 \
actor_rollout_ref.actor.fsdp_config.param_offload=true \
actor_rollout_ref.actor.fsdp_config.grad_offload=true \
actor_rollout_ref.actor.fsdp_config.optimizer_offload=true \
actor_rollout_ref.rollout.log_prob_micro_batch_size=128 \
actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
actor_rollout_ref.rollout.name=vllm \
actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
actor_rollout_ref.ref.log_prob_micro_batch_size=128 \
actor_rollout_ref.ref.fsdp_config.param_offload=True \
actor_rollout_ref.actor.kl_loss_coef=0.001 \
actor_rollout_ref.actor.kl_loss_type=low_var_kl \
algorithm.no_think_rl=false \
actor_rollout_ref.rollout.n_agent=5 \
actor_rollout_ref.rollout.temperature=1 \
actor_rollout_ref.actor.state_masking=true \
trainer.logger=['wandb'] \
+trainer.val_only=false \
+trainer.val_before_train=true \
trainer.default_hdfs_dir=null \
trainer.n_gpus_per_node=8 \
trainer.nnodes=1 \
trainer.save_freq=100 \
trainer.test_freq=50 \
trainer.project_name=$WAND_PROJECT \
trainer.experiment_name=$EXPERIMENT_NAME \
trainer.total_epochs=15 \
trainer.total_training_steps=305 \
trainer.default_hdfs_dir=null \
trainer.default_local_dir=verl_checkpoints/$EXPERIMENT_NAME \
max_turns=4 \
retriever.url="http://127.0.0.1:8000/retrieve" \
retriever.topk=3 \
2>&1 | tee $EXPERIMENT_NAME.log

92
scripts/nq_hotpotqa/train_ppo.sh Normal file
View File

@@ -0,0 +1,92 @@
data_name=nq_hotpotqa_train
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export DATA_DIR=data/${data_name} # first download the data from https://huggingface.co/datasets/PeterJinGo/nq_hotpotqa_train
WAND_PROJECT="Search-R1"
export BASE_MODEL='meta-llama/Llama-3.2-3B'
export EXPERIMENT_NAME=${data_name}-search-r1-ppo-llama3.2-3b-em
# export BASE_MODEL='meta-llama/Llama-3.2-3B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-llama3.2-3b-it-em
# export BASE_MODEL='meta-llama/Llama-3.1-8B'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-llama3.1-8b-em
# export BASE_MODEL='meta-llama/Llama-3.1-8B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-llama3.1-8b-it-em
# export BASE_MODEL='Qwen/Qwen2.5-3B'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-qwen2.5-3b-em
# export BASE_MODEL='Qwen/Qwen2.5-3B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-qwen2.5-3b-it-em
# export BASE_MODEL='Qwen/Qwen2.5-7B'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-qwen2.5-7b-em
# export BASE_MODEL='Qwen/Qwen2.5-7B-Instruct'
# export EXPERIMENT_NAME=${data_name}-search-r1-ppo-qwen2.5-7b-it-em
# set -x
export VLLM_ATTENTION_BACKEND=XFORMERS # vllm + qwen2-7b with flash_attn has some issues
# max_prompt_length = (config['training']['max_start_length'] + config['training']['max_response_length'] * (config['training']['max_turns'] - 1) + config['training']['max_obs_length'] * config['training']['max_turns'])
PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
data.train_files=$DATA_DIR/train.parquet \
data.val_files=$DATA_DIR/test.parquet \
data.train_data_num=null \
data.val_data_num=null \
data.train_batch_size=512 \
data.val_batch_size=256 \
data.max_prompt_length=4096 \
data.max_response_length=500 \
data.max_start_length=2048 \
data.max_obs_length=500 \
data.shuffle_train_dataloader=True \
algorithm.adv_estimator=gae \
actor_rollout_ref.model.path=$BASE_MODEL \
actor_rollout_ref.actor.optim.lr=1e-6 \
actor_rollout_ref.model.enable_gradient_checkpointing=true \
actor_rollout_ref.model.use_remove_padding=True \
actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=0.95 \
actor_rollout_ref.actor.ppo_mini_batch_size=256 \
actor_rollout_ref.actor.ppo_micro_batch_size=64 \
actor_rollout_ref.actor.fsdp_config.param_offload=true \
actor_rollout_ref.actor.fsdp_config.grad_offload=true \
actor_rollout_ref.actor.fsdp_config.optimizer_offload=true \
actor_rollout_ref.rollout.log_prob_micro_batch_size=128 \
actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
actor_rollout_ref.rollout.name=vllm \
actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
actor_rollout_ref.ref.log_prob_micro_batch_size=128 \
actor_rollout_ref.ref.fsdp_config.param_offload=True \
actor_rollout_ref.rollout.n_agent=1 \
actor_rollout_ref.rollout.temperature=1 \
actor_rollout_ref.actor.state_masking=true \
critic.optim.lr=1e-5 \
critic.model.use_remove_padding=True \
critic.optim.lr_warmup_steps_ratio=0.05 \
critic.model.path=$BASE_MODEL \
critic.model.enable_gradient_checkpointing=true \
critic.ppo_micro_batch_size=8 \
critic.model.fsdp_config.param_offload=true \
critic.model.fsdp_config.grad_offload=true \
critic.model.fsdp_config.optimizer_offload=true \
algorithm.kl_ctrl.kl_coef=0.001 \
algorithm.no_think_rl=false \
trainer.critic_warmup=0 \
trainer.logger=['wandb'] \
+trainer.val_only=false \
+trainer.val_before_train=true \
trainer.default_hdfs_dir=null \
trainer.n_gpus_per_node=8 \
trainer.nnodes=1 \
trainer.save_freq=100 \
trainer.test_freq=50 \
trainer.project_name=$WAND_PROJECT \
trainer.experiment_name=$EXPERIMENT_NAME \
trainer.total_epochs=15 \
trainer.total_training_steps=305 \
trainer.default_hdfs_dir=null \
trainer.default_local_dir=verl_checkpoints/$EXPERIMENT_NAME \
max_turns=4 \
retriever.url="http://127.0.0.1:8000/retrieve" \
retriever.topk=3 \
2>&1 | tee $EXPERIMENT_NAME.log