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: 10000 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-10000

Warning

Seed-43 resume run from checkpoint-5000 without loading rng_state_*.pth. This run showed repeated narrow train accuracy drops; checkpoint-10000 is near a warning-level eval copy-accuracy drop and should be treated as diagnostic.

Metrics At Checkpoint

Metric	Value
eval/loss	0.306291
eval/mean_copy_accuracy	0.972965
eval/mean_gen_accuracy	0.895226
eval/mean_token_accuracy	0.903894
eval/z_loss	0.00548837
eval/copy_logits_max	-14.2332
train/loss at step 9999	0.2906
train/mean_copy_accuracy at step 9999	0.967167
train/mean_gen_accuracy at step 9999	0.891601
train/mean_token_accuracy at step 9999	0.908741

File Contents

This repo excludes global_step*, rng_state*.pth, scheduler.pt, trainer_state.json, training_args.bin, and zero_to_fp32.py.

Accuracy Trace

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