reasoning-gym/training/README.md

Reasoning Gym Model Training

Training codebase for training LLMs using Reasoning Gym procedural dataset generators.

Note: qwen-math/ directory contains the code from the Tina project, used for the Qwen2.5 3B RG-Math training. This is separate from the rest of our training/evaluation codebase.

This readme documents:

• Training environment setup and usage example
• Converting training checkpoints to HuggingFace format
• Evaluation setup and usage for eval on RG data
• Evaluation setup and usage for external benchmarks

Requirements

1. Prepare and activate a Python 3.11 virtual environment however you prefer.
2. Install Reasoning Gym:

cd reasoning-gym/
pip install -e .

3. Install training-specific Python package dependencies:

pip install ray wandb
pip install torch==2.6.0

4. Install veRL (tested with HEAD c34206925e2a50fd452e474db857b4d488f8602d):

pip install git+https://github.com/volcengine/verl.git@c6dc8b73cf011aa75b8c6a47b0322f50aed800ad#egg=verl

5. Install vLLM:

pip install vllm==0.6.3 transformers==4.50.3 fire==0.7.0

6. Install flash attention

pip install flash-attn --no-build-isolation

6. Log in to HF and W&B:

huggingface-cli login
wandb login

Usage

First, activate the virtual environment you prepared.

Example GRPO training usage:

python3 -u train_grpo.py --config-name llama3.1_1b_grpo \
    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
    trainer.project_name=rg-test \
    trainer.experiment_name=verl_grpo_llama3.1_1b \
    trainer.n_gpus_per_node=2 $@ 2>&1 | tee verl_output.log

Then, having saved this as a bash script such as train.sh, run it:

CUDA_VISIBLE_DEVICES=0,1 bash train.sh

CUDA_VISIBLE_DEVICES is set to 0,1 to use the first two GPUs on the machine (see nvidia-smi output). This can be adjusted as needed. tensor_model_parallel_size and n_gpus_per_node should also be set to the number of GPUs you are using.

You can change all configuration options by either modifying the config YAML (in this case, config/llama3.1_1b_grpo.yaml) or providing them as arguments to the Python script. Note that the batch sizes set in the Llama 1B and Qwen 1.5B configs are as high as it was possible for me to set them for the puzzles dataset mix on 2xA6000 GPUs without OOMs. Depending on the hardware you use and the datasets you train on, you may need to adjust these.

Exporting from FSDP checkpoint to HF model checkpoint

After training your model the weights are saved across as a sharded checkpoints across several files. To faciliate simple evaluation of your trained model you may want to convert this into a HF model checkpoint. We have added a utility script to convert your sharded checkpoint into a hf checkpoint.

To run this script. Navigate to the training directory and run the following

python load_fsdp_to_hf.py /path/to/fsdp/checkpoint/global_step_num/actor /path/to/hugginface/checkpoint/global_step_num/actor/huggingface saved_model_name

For example

python utils/load_fsdp_to_hf.py checkpoints/rg-test/intra_reasoning_algorithmic_qwen_3b_composite/global_step_400/actor/ checkpoints/rg-test/intra_reasoning_algorithmic_qwen_3b_composite/global_step_400/actor/huggingface qwen3b

Run evaluations

From here you may to run evaluations of your trained model. In the training/evaluation directory there is a script evaluate_model.py which you csn run to evaluate your trained model on a specific dataset. You specify evaluation parameters in a yaml file. This evaluation can point to either a local or remote model. For example the configuration file training/evaluation/eval_algorithmic_composite.yaml specifies the path to a local model which is stored as a hugginface checkpoint at training/utils/qwen3b_500 (note that you have to convert to fsdp checkpoint to hf checkpoint for evaluation script to work as shown in the previous step).

Run the script

export VLLM_ATTENTION_BACKEND=XFORMERS

Navigate to evaluations directory:

python evaluate_model.py --config path-to-yaml

For example

python evaluate_model.py --config eval_algorithmic_composite.yaml

External benchmark evaluations

We additionally evaluate some models on external benchmarks using the Language Model Evaluation Harness from Eleuther AI.

We utilise the llama branch for the Llama 3 MATH and GSM8K evaluation configurations it provides, for the fairest possible comparison against Meta's original Llama 3 model.

git clone https://github.com/EleutherAI/lm-evaluation-harness.git
cd lm-evaluation-harness
git checkout llama
pip install -e .

For our Llama 3 3B RG-Math model, we evaluate both the original model and ours by directly using the Llama 3 configs provided by LMEH:

# tasks used: llama_math, gsm8k_cot_llama
lm_eval --model vllm --model_args pretrained=/path/to/model --tasks llama_math --batch_size auto --output_path results/ --apply_chat_template --fewshot_as_multiturn

For our Qwen2.5 3B RG-Math model, we evaluate using a tweaked version of the same task configs. The system prompt used in RL is also used in evaluation for the RG-Math model. The original Qwen2.5 model was tested with the same system prompt, but performed worse than with the standard CoT prompt, so the final evaluation score utilised the standard prompt.

# tasks used: llama_math (edited, see below), gsm8k_cot_rg
lm_eval --model vllm --model_args pretrained=/path/to/model --tasks llama_math --batch_size auto --output_path results/

The RG-specific task configs for LMEH are contained in training/evaluations/lmeh/ in this repository. To run the llama_math eval, replace llama_math_algebra in the relevant LMEH tasks directory with the RG one provided.