README.md

SkillZero Best Checkpoints Export

This package contains the two best checkpoints selected by validation success rate from the completed SkillZero runs.

Included Checkpoints

1. ALFWorld global_step_160

   • Validation metric: val/success_rate = 0.594
   • Archive path: checkpoints/SkillZero_alfworld/skillzero_alfworld_vl_3b_safe/global_step_160

2. ALFWorld global_step_150

   • Validation metric: val/success_rate = 0.477
   • Archive path: checkpoints/SkillZero_alfworld/skillzero_alfworld_vl_3b_safe/global_step_150

Related Search Checkpoint

The best Search checkpoint by validation success rate is not included in the "top two overall" package, but is useful for reproducing the Search run:

• Search global_step_180
• Validation metric: val/success_rate = 0.356
• Test metrics:
  • test/full_skill/success_rate = 0.282
  • test/no_skill/success_rate = 0.310
• Checkpoint path, if packaged separately: checkpoints/SkillZero_search/skillzero_search_vl_3b_local_retriever/global_step_180

Hardware Used

Training was submitted through Slurm on the a100 partition.

• ALFWorld:

  • GPUs: 4 x A100
  • CPUs per task: 32
  • Memory: 200GB
  • Time limit: 2 days

• Search local retriever:

  • GPUs: 4 x A100 allocated
  • Training used GPUs 0,1,2
  • Local retriever used GPU 3
  • CPUs per task: 32
  • Memory: 220GB
  • Time limit: 2 days

Runtime Notes

• Python environment name used on the cluster: skillzero
• Retriever environment name: retriever
• Main model: Qwen/Qwen2.5-VL-3B-Instruct
• Training entry point: python3 -m verl.trainer.main_ppo
• Original training logs are not required to use the checkpoints.

Restore

After extracting the archive, place checkpoint directories under:

checkpoints/SkillZero_alfworld/skillzero_alfworld_vl_3b_safe/

Then use trainer.resume_mode=resume_path and set trainer.resume_from_path to the target global_step_* directory.