CoVT-Phase2-3expert-Strict

Canonical strict 3-expert Phase2 LoRA + merged weights for our CoVT reproduction. Trained after the model.global_steps-on-resume bug was fixed, so the anchor-loss schedule matches the paper.

Contents

Path	Description
adapter_config.json, adapter_model.safetensors, non_lora_state_dict.bin	Final Phase2 LoRA adapter + non-LoRA state.
checkpoint-9000/, checkpoint-10000/	Intermediate Phase2 checkpoints (LoRA + DeepSpeed Zero-2 optimizer states).
stage4_merged_strict/	Final adapter merged into Qwen2.5-VL-7B base — ready for inference / eval.
eval_stage4_strict/	VLMEvalKit results on CV-Bench 2D / 3D.
config.json, config_patched.py, trainer_patched.py, trainer_state.json	Final config and trainer state.
runs/	Tensorboard event files.
*.log	Phase2 strict training / eval / continuation logs.

Companion repos

• Steven668866/CoVT-3expert-Stage1-LoRA — Phase1 6K-step LoRA adapter (clean).
• Steven668866/CoVT-Phase2-3expert-Full — Historical repo from the non-strict run; buggy adapter weights have been removed (2026-06-29). Only the clean Phase1 merge + scripts + logs remain there.

Caveat

This is a 3-expert variant of CoVT. The paper's main configuration uses 4 experts, so absolute numbers are not directly comparable to the paper's headline table.