suss-models / MODELS.md
nonberew's picture
SUSS checkpoints (v0.1.3), model card and manifest
eaca32f verified
|
Raw History Blame Contribute Delete
6.29 kB

SUSS trained models

All checkpoints used in the paper. Load with torch.load(path, map_location='cpu'); GPT checkpoints hold cfg (a suss.model.GPTConfig as a dict) and state (the state dict). See notebooks/quickstart.ipynb.

file what parameters MB
runs/a/seed0.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed1.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed2.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed3.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed4.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed5.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed6.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed7.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed8.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/a/seed9.pt A: in-context predictor on linear systems (n=2, m=1), 20k steps; seeds 0-2 pre-registered, 3-9 replication 800,258 3.2
runs/b/block_tf_seed0.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/block_tf_seed1.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/block_tf_seed2.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/bouncer_tf_seed0.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/bouncer_tf_seed1.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/bouncer_tf_seed2.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,258 3.2
runs/b/pendulum_tf_seed0.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,515 3.2
runs/b/pendulum_tf_seed1.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,515 3.2
runs/b/pendulum_tf_seed2.pt B: in-context predictor on MuJoCo pendulum / block / bouncer (includes token normalisation stats) 800,515 3.2
runs/b/block_mlp_seed0.pt B: no-context MLP baseline 67,330 0.3
runs/b/block_mlp_seed1.pt B: no-context MLP baseline 67,330 0.3
runs/b/block_mlp_seed2.pt B: no-context MLP baseline 67,330 0.3
runs/b/bouncer_mlp_seed0.pt B: no-context MLP baseline 67,330 0.3
runs/b/bouncer_mlp_seed1.pt B: no-context MLP baseline 67,330 0.3
runs/b/bouncer_mlp_seed2.pt B: no-context MLP baseline 67,330 0.3
runs/b/pendulum_mlp_seed0.pt B: no-context MLP baseline 67,843 0.3
runs/b/pendulum_mlp_seed1.pt B: no-context MLP baseline 67,843 0.3
runs/b/pendulum_mlp_seed2.pt B: no-context MLP baseline 67,843 0.3
runs/c/bc_seed0.pt C: controller after the DAgger warm start (imitation only) 800,129 3.2
runs/c/bc_seed1.pt C: controller after the DAgger warm start (imitation only) 800,129 3.2
runs/c/bc_seed2.pt C: controller after the DAgger warm start (imitation only) 800,129 3.2
runs/c/seed0.pt C: final controller (warm start + log-cost fine-tuning, validation-selected checkpoint) 800,129 3.2
runs/c/seed1.pt C: final controller (warm start + log-cost fine-tuning, validation-selected checkpoint) 800,129 3.2
runs/c/seed2.pt C: final controller (warm start + log-cost fine-tuning, validation-selected checkpoint) 800,129 3.2
runs/c/seed0_last.pt C: last fine-tuning checkpoint (not selected) 800,129 3.2
runs/c/seed1_last.pt C: last fine-tuning checkpoint (not selected) 800,129 3.2
runs/c/seed2_last.pt C: last fine-tuning checkpoint (not selected) 800,129 3.2
runs/c/seed0_raw.pt C: pilot, warm start + raw cost (diverged; selected = warm start) 800,129 3.2
runs/c/seed0_raw_last.pt C: pilot, warm start + raw cost (diverged; selected = warm start) 800,129 3.2
runs/c/seed0_cvar.pt C: pilot, warm start + CVaR (diverged; selected = warm start) 800,129 3.2
runs/c/seed0_cvar_last.pt C: pilot, warm start + CVaR (diverged; selected = warm start) 800,129 3.2
runs/c/fixed_seed0.pt C: best constant-gain baseline (a 1x2 gain matrix) 2 0.0
runs/c/fixed_seed1.pt C: best constant-gain baseline (a 1x2 gain matrix) 2 0.0
runs/c/fixed_seed2.pt C: best constant-gain baseline (a 1x2 gain matrix) 2 0.0
runs/scale/deep8.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 1,593,346 6.4
runs/scale/long60k.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 800,258 3.2
runs/scale/n4m2_s0.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 800,900 3.2
runs/scale/n4m2_s1.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 800,900 3.2
runs/scale/n4m2_s2.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 800,900 3.2
runs/scale/small64.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 103,554 0.4
runs/scale/wide256.pt Scale study: long60k, deep8, wide256, small64, n4m2_s{0,1,2} 3,173,378 12.7