How om-ai-lab/vlm-r1 works
Main Components Identification
• Core Training Module: The heart of the system lies in the “src/open-r1-multimodal” directory. It includes:
– Setup scripts (setup.sh, Dockerfile)
– Configuration files (YAML configs in “configs” and “data_config” folders)
– Model code under “src/open_r1” with sub-modules for evaluation, generation, different GRPO variants, SFT, and additional trainer modules.
• Data Handling & Preprocessing: There are directories and scripts for data management:
– “src/open-r1-multimodal/data_config” for dataset configurations.
– “src/open-r1-multimodal/data_jsonl” and dedicated scripts like “grpo_jsonl.py” to handle custom JSONL inputs and multi-image input formats.
• Run & Training Scripts: Located in “src/open-r1-multimodal/run_scripts” and “local_scripts”, these take care of launching different training modes (e.g., single-node, multi-node, GUI-based evaluation) as well as fine-tuning strategies.
• Evaluation Module: Found in “src/eval”—includes evaluation scripts (e.g., test_rec_r1.py and test_rec_baseline.py) and setups for measuring performance.
Relationships and Interactions
• The training scripts and run scripts (inside “run_scripts” and “local_scripts”) interact with core modules (in “src/open_r1” and “trainer”) by passing configuration parameters (from YAML and JSON config files).
• Data flows from external sources (images and JSON/JSONL annotation files) into data ingestion routines (as configured by “data_config/rec.yaml”) and then gets processed by the GRPO or SFT training pipelines.
• The evaluation scripts in “src/eval” depend on the output models generated by the training process and feed into external performance reporting tools (e.g., WandB, as hinted in the training commands).
Architectural Patterns and Design Principles
• Modular Structure: The code base is split into clearly delineated modules (e.g., model configuration, training, utilities, and evaluation), which helps isolate concerns.
• Configuration-driven Execution: Extensive use of YAML files for model, training, and data configuration promotes flexibility and dynamic adjustment.
• Support for Distributed Training: The use of “torchrun”, multi-node training scripts, and Deepspeed configuration (e.g., through “zero2.yaml”, “zero3.yaml” scripts) indicates a design aimed at scaling across multiple GPUs/nodes.
• Fine-tuning Approaches: Separation between full fine-tuning (GRPO) and parameter-efficient fine-tuning (LoRA, SFT via LLaMA-Factory) points to extensible design for experimental research in model training.
Relevant Technologies, Frameworks, or Libraries
• Python (with a clear version requirement – Python 3.10)
• Deep learning frameworks: PyTorch, Deepspeed, and torchrun for distributed training.
• Docker for containerized deployment.
• Additional integration with external libraries such as WandB for logging and monitoring.
• LLaMA-Factory integration for SFT model training.
Overall Detailed Approach
• Nodes for auxiliary scripts (such as multi-node training scripts and local scripts for creating specific training data or launching experiments).
• Don’t oversimplify—detail each component such that interactions like model generation, evaluation callbacks, and configuration dependencies are clearly visible.