INTACT / INTACT_GOAL_E1_SINGLE_README.md
jw8866's picture
release: publish Goal E1 single-task checkpoints
761d24b verified
|
Raw
History Blame Contribute Delete
1.47 kB

INTACT Goal E1 Single-Task Four-Slot 7/7

This release contains 12 independently trained checkpoints: four tasks for each training seed (0, 42, and 3072). Each archive follows the repository convention of grouping four task shards by training seed.

Archives

Archive Bytes SHA-256
intact-goal-e1-single-seed0.tar.gz 310459876 4fc8cd34768ddc431cf0138d8016a1e65330ae1cd5bd060f0846cebe39c0611d
intact-goal-e1-single-seed42.tar.gz 310468493 b5535232e79e7672832dc824c4b19111c18f4fd43f692f3db49aff0fb562343c
intact-goal-e1-single-seed3072.tar.gz 310468271 c1aa0f5427e16b553efca69bb9483e688dca0723920f358c0e7080b9ad2fa541

Each archive contains checkpoints/intact_goal_<task>_s<seed>_p77_v1/weights_epoch_1.pt for PushT, Cube, Reacher, and TwoRoom. Verify archives with INTACT_GOAL_E1_SINGLE_SHA256SUMS and extracted shards with INTACT_GOAL_E1_SINGLE_MANIFEST.json.

Contract

  • Original Goal INTACT, independently trained per task
  • Four-slot Actor, 7/7 window, 8 frames, terminal goal z7
  • One epoch, batch 256, workers 6, BF16 mixed precision, Math SDPA
  • Train seeds 0/42/3072; Official Direct evaluation seeds 0/1/42, 100 episodes per cell
  • 12 checkpoints and 36 evaluations passed; overall success rate 93.72% (3374/3600)

These checkpoints use the clean four-slot runtime. They are intentionally incompatible with the legacy five-slot Actor grammar documented for the paper E5 archives.