This is a cleaned reproduction workspace for the official DyME project, focused on the ChartQA Table 2 style Full DyME / Full Visual Supervision experiment with LLaVA-OV-S.
Scope
Official DyME source: code/DyME
Official commit: bf4ca908c290030526be41344b3f0ec707ef9bb0
Base model: LLaVA-OV-S, llava-onevision-qwen2-0.5b-ov-hf
Task: ChartQA train/test
Training data: runtime/chartqa/train_new_prerefine.json
Online visual supervision: enabled through RewardCalculator and ContextRefiner
Helper LLM: Qwen/Qwen2.5-14B-Instruct-AWQ, served by an OpenAI-compatible lmdeploy API
Same-machine layout: one 8-GPU node, GPU 7 for helper, GPUs 0-6 for DyME training/evaluation
This package includes small runtime safety fixes for online helper failures and a grpo route alias; see metadata/debug_20260724.md.
The official config uses 8 training processes and 10 epochs. On a single 8-card machine where the helper must run locally, this package defaults to 7 training processes plus one helper GPU, while preserving the official DyME hyperparameters and online visual supervision path.
For a tiny one-step model-training smoke, add RUN_TRAIN_SMOKE=1. This uses a fake OpenAI-compatible helper and is only intended to validate code paths.
What Is Not Included
Large local-only artifacts are intentionally excluded from upload and cleanup targets:
conda/env directories
interrupted or non-final run checkpoints
wandb/offline logs
SLAKE/BiomedGPT experiments
helper logs and pid files
model weights
ChartQA image files
If this repo is moved to a new machine, update DYME_PRETRAINED_MODEL_PATH, HF_HOME, and the image paths inside runtime/chartqa/train_new_prerefine.json or regenerate the preprocessed ChartQA JSON.
