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: 17000 Source path on launch pod:

/home/irteam/data/junhyeok/cache/tvc_Qwen2_5-VL-7B-Instruct_v2_pgn_maximg672_bs4_ga4_resume5000_seed43_norng_h200x4_save500_max24465/checkpoint-17000

Warning

Unstable checkpoint. The seed-43 resume run showed several severe train-side accuracy drops around the 17000 window; this artifact is uploaded for recovery and analysis, not as a recommended final model.

After checkpoint-17000, W&B summary reached train/global_step 17021 with train/mean_copy_accuracy 0.9315, train/mean_gen_accuracy 0.9031, train/loss 0.2918, and train/z_loss 0.0183 before the run was stopped.

Metrics At Checkpoint

Metric	Value
eval/loss	0.215967
eval/mean_copy_accuracy	0.983707
eval/mean_gen_accuracy	0.925558
eval/mean_token_accuracy	0.932245
eval/z_loss	0.00890145
eval/copy_logits_max	-27.4164
train/loss at step 16999	0.2439
train/mean_copy_accuracy at step 16999	0.973012
train/mean_gen_accuracy at step 16999	0.90822
train/mean_token_accuracy at step 16999	0.925382

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.