2bidoubi's picture
Model card: corrected Items hand-over scores, full fine-tune 78/100 at step 15000 recommended
18489c8 verified
|
Raw History Blame Contribute Delete
1.12 kB

items_handover

Checkpoints for the RoboSynChallenge task items_handover, stored as <run>/<step>/ openpi checkpoint folders (params/, assets/<repo_id>/norm_stats.json, train_state/ where uploaded). Results and recipes are in the repository README.

  • pi05_lora_v1: steps 2500, 5000, 7500, 10000, 12500, 14999 (official data; best 38/100 at step 12500)
  • pi05_full_v1: steps 5000, 10000, 15000, 19999 (full fine-tune on the official data; 58, 72, 78 and 77 per 100; step 15000 is the recommended Items hand-over checkpoint)
  • pi05_short_v1: steps 1000, 2000, 3000, 4000, 4999 (5000-step full fine-tune on the official data, lr 1e-5; 23, 34, 44, 59 and 65 per 100)
  • pi05_mix_lora_v1: steps 2500, 5000, 7500, 10000, 12500, 14999 (official + team-generated data; 16/100 at step 12500)
  • pi05_aug_lora_v1: steps 2500, 5000, 7500, 10000, 12500, 14999 (team-generated data only; 0 successes, kept for the record)

Scores of the full and short full fine-tunes and of the mixed-data LoRA were re-measured on 2026-10-11 after a broken success check in our local task code was fixed (see the repository README).