Skip to content
ExplainableMLPublic

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 

Repository files navigation

VERVE: Verified Visual Search

Code for VERVE: Verified Visual Search.

Training

Training lives in EasyR1/, a fork of hiyouga/EasyR1 (GRPO on top of veRL) at upstream commit dd71bbd plus PR #11001 for Qwen3.5. One tree trains every backbone:

MODEL Backbone
qwen25vl Qwen/Qwen2.5-VL-7B-Instruct
qwen3vl Qwen/Qwen3-VL-8B-Instruct
qwen35 Qwen/Qwen3.5-9B
mimo XiaomiMiMo/MiMo-VL-7B-SFT-2508

and both modes:

MODE Reward Prompt
plain examples/reward_function/perception_multinode.py examples/format_prompt/perception.jinja
opsv examples/reward_function/perception_verify.py examples/format_prompt/perception_verify2_1.jinja

Setup

cd EasyR1
pip install -r requirements.txt     # vllm 0.18, torch 2.10
pip install transformers==5.5.3     # after vllm, which declares transformers<5

This is the environment we tested (Python 3.12). verl is imported from the EasyR1/ directory, so there is nothing else to install.

Download VERVE-Data and extract the image archives as its card describes. The policy and judge models must be in the local Hugging Face cache (HF_HOME).

Run

train_verve.sh is a SLURM launcher for 7 nodes with 4 GPUs each: 6 nodes train, 1 node serves the reward judge (Qwen/Qwen3-30B-A3B-Instruct-2507). Set your account and partition in its header, then:

cd EasyR1
MODEL=qwen3vl MODE=opsv \
TRAIN_PARQUET=/data/VERVE-Data/train_s57_opsv.parquet \
IMAGE_DIR=/data/VERVE-Data/images \
HF_HOME=/path/to/hf_cache \
sbatch train_verve.sh

Use train_s57_plain.parquet with MODE=plain. The shared cluster logic (Ray cluster, judge servers, resume across the walltime limit) is in train_common.sh; every hyperparameter is a config override there or in examples/perception_config.yaml.

ANNEAL="START END" (OPSV only) retires the verification terms of the reward linearly between the two steps, so training ends on the plain accuracy reward. It defaults to "40 80" for qwen35 and is off for the other backbones, which matches the released checkpoints.

Merging a checkpoint

python scripts/model_merger.py --local_dir verl_exp/<run>/global_step_<N>/actor

For Qwen3.5, also run python scripts/fix_qwen35_merged_keys.py --hf_dir <...>/actor/huggingface afterwards.

Acknowledgements

Built on EasyR1 and veRL; the training recipe starts from Zooming without Zooming.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages