Source run: tvc_Qwen2_5-VL-7B-Instruct_v2_pgn_maximg672_bs4_ga4_resume5000_seed43_norng_h200x4_save500_max24465 W&B: https://wandb.ai/mir_lab/mm-pgn/runs/p08ehggd Checkpoint step: 15000 Source path on launch pod:

/home/irteam/data/junhyeok/cache_backup/tvc_Qwen2_5-VL-7B-Instruct_v2_pgn_maximg672_bs4_ga4_resume5000_seed43_norng_h200x4_save500_max24465/checkpoint-15000

Warning

Seed-43 resume run from checkpoint-5000 without loading rng_state_*.pth. This checkpoint has a cleaner eval than checkpoint-10000/17000, but the same run later showed repeated train accuracy drops.

Metrics At Checkpoint

Metric	Value
eval/loss	0.215134
eval/mean_copy_accuracy	0.994089
eval/mean_gen_accuracy	0.921149
eval/mean_token_accuracy	0.929532
eval/z_loss	0.00174698
eval/copy_logits_max	-64.722
train/loss at step 14999	0.22
train/mean_copy_accuracy at step 14999	0.994246
train/mean_gen_accuracy at step 14999	0.905995
train/mean_token_accuracy at step 14999	0.924563

File Contents

This repo includes model weights, tokenizer/processor files, Trainer state, scheduler state, RNG states, and DeepSpeed optimizer/model-state shards under global_step*/.

Accuracy Trace

The seed-43 run-level accuracy plot is included as accuracy_plot.png.