Skip to content

jy1993/SimpleRL

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
assets/images
assets/images
 
 
cold_start
cold_start
 
 
data
data
 
 
scripts
scripts
 
 
Readme.md
Readme.md
 
 
agent_eval.py
agent_eval.py
 
 
agent_utils.py
agent_utils.py
 
 
args.py
args.py
 
 
data_utils.py
data_utils.py
 
 
grpo_utils.py
grpo_utils.py
 
 
llm_utils.py
llm_utils.py
 
 
main.py
main.py
 
 
prompt.py
prompt.py
 
 
requirements.txt
requirements.txt
 
 

Repository files navigation

SimpleRL

My implementations of RL algorithms like GRPO/GSPO with minimal code.

RL Infrastructure

rl_infra

Features

  • Supported models
    • Qwen2/Qwen2.5/Qwen3 language models
    • Qwen2.5 vision language models
  • Supported algorithms
    • GRPO
    • Dr-GRPO
    • GSPO
    • KL-Conv
    • StableReinforce
  • Supported tricks
    • clip higher from DAPO
    • token level policy loss
    • dual clip
    • kl term removal

Requirements

pip install -r requirements

Train on dapo_math

bash scripts/run_agent_math.sh

Dataset

https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k

Results on math benchmarks(metric:avg@32/pass@32)

Model MATH500 AIME24 AIME25 HMMT Feb.25 BeyondAIME AVG
Qwen2.5-7B-Instruct + sft 0.619/0.884 0.105/0.5 0.065/0.4 0.047/0.367 0.028/0.25 0.173/0.48
Qwen2.5-7B-Instruct + rl 0.718/0.91 0.261/0.6 0.227/0.533 0.125/0.367 0.068/0.33 0.28/0.548
Qwen3-4B-Instruct + sft 0.777/0.916 0.268/0.7 0.202/0.333 0.130/0.5 0.109/0.44 0.297/0.578
Qwen3-4B-Instruct + rl 0.843/0.928 0.442/0.8 0.385/0.733 0.239/0.6 0.209/0.52 0.424/0.716

Training curves:

train_curves_of_qwen2.5_7b_instruct_on_dapo

train_curves_of_qwen2.5_7b_instruct_on_dapo

train_curves_of_qwen3_4b_instruct_on_dapo

train_curves_of_qwen3_4b_instruct_on_dapo

Reasoning Examples

agent_math_response_aime_0 agent_math_response_aime_1

Train on logic

bash scripts/run_logic.sh

Dataset

https://github.com/Unakar/Logic-RL/tree/main/data/kk/instruct

Results on test set

Size Algorithm Bits LR KL Group Size Steps Test Score
3B GRPO AMP 1e-6 0 8 1600 0.12->0.54
7B GRPO AMP 1e-6 0 8 1350 0.23->0.89

Details

Model 2ppl 3ppl 4ppl 5ppl 6ppl 7ppl 8ppl
Qwen2.5-3B-Instruct 0.37 0.13 0.17 0.12 0.04 0.02 0.02
Qwen2.5-3B-Instruct-GRPO 0.76 0.70 0.68 0.50 0.47 0.33 0.33
Qwen2.5-7B-Instruct 0.56 0.35 0.23 0.25 0.14 0.09 0.02
Qwen2.5-7B-Instruct-GRPO 0.97 0.96 0.96 0.94 0.88 0.79 0.72

Training curves:

train_curves_of_qwen2.5_3B_instruct_grpo

train_curves_of_qwen2.5_3b_instruct_on_logic

train_curves_of_qwen2.5_7B_instruct_grpo

train_curves_of_qwen2.5_7b_instruct_on_logic

Train on geometry3k

bash scripts/run_geometry3k.sh

dataset

https://huggingface.co/datasets/hiyouga/geometry3k

Results on test set

Size Algorithm Bits LR KL Group Size Steps Test Score
3B GRPO AMP 1e-6 0 8 700 0.24->0.43
3B GSPO AMP 1e-6 0 8 750 0.24->0.43
3B stable reinforce AMP 1e-6 0 12 1200 0.25->0.44
3B kl-conv AMP 1e-6 0 12 900 0.23->0.45
7B GRPO AMP 1e-6 0 8 800 0.38->0.50

Training curves:

train_curves_of_qwen2.5_3B_vl_instruct_grpo

train_curves_of_qwen2.5_3b_vl_instruct_on_geometry3k

train_cuvres_of_qwen2.5_7B_vl_instruct_grpo

train_curves_of_qwen2.5_7b_vl_instruct_on_geometry3k

TODO

  • train on math
  • train with Dr-GRPO/GSPO/KL-Conv/StableReinforce algos
  • support dynamic sampling from dapo
  • support ppo/reinforce++/RLOO
  • support vision language models
  • support Retrieval-Augmented Reasoning
  • support agent training
  • support code eval
  • support vllm infer parallelism(dp2, tp4)
  • add ulysses sequence parallelism
  • support padding free training

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

188 stars

Watchers

16 watching

Forks

20 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages