cvis-tmu/qwen2_5vl-7b-lora-sft-CoT_traineval_1epochs_traininglog

Qwen2.5-VL-7B LoRA adapter. Trained for 1 epoch (620 steps) on 4x H100 GPUs (job: 428685 on TamIA). Dataset: Scene30k + SpatialSSRL_coldstart + 3DThinker10k (mix_strategy=concat). It is identical to qwen2_5vl-7b-lora-sft-CoT_traineval_1epochs, but here we log batch predictions every 10 training steps and log the evaluation subset after 1 epoch.

initial commit