grail_v1_fcur_b64_sft_e3_ck300

grail_v2op_nograph_b64_sft_e3_ck640

grail_v2op_trajqsp_b64_sft_e3_ck640

grail_v2op_fcursp_b64_sft_e3_ck640

grail_v2op_trajq_b64_sft_e3_ck640

oracle_reason_s2_b64_sft_e3_ck300

oracle_answer_b64_sft_e3_ck300

grail_v2op_traj_b64_sft_e3_ck640

grail_v1_traj_b64_sft_e3_ck300

grail_v2op_fcur_lora64_b64_sft_e3_merged

grail_v2op_fcur_frombase_b64_sft_e0_ck640

grail_v2op_fcurlong_b64_sft_e3_ck640

grail_v2op_fcur_sr_b64_sft_e3_ck640

grail_v2op_fcur_long_b64_sft_e3_ck1280

grail_v5_fcur_op48k_b64_sft_e3_ck800

grail_v2op_fcur_b64_sft_e3_ck640

grail_v1_trajq_rtx_b64_sft_e3_ck300

grail_v3c_343_b64_sft_e3_ck640

grail_v2opv_343_b64_sft_e3_ck640

grail_v2op_343_b64_sft_e3_ck640

grail_v1_343_b64_sft_e3_ck300

her_nemi_graph_fcur_rw_b64_sft_e3_ck300

her_nemi_graph_fcur_geo_b64_sft_e3_ck300

her_nemi_graph_343bal_geo_b64_sft_e3_ck300

her_nemi_graph_343_geo_b64_sft_e3_ck300

rename her_nemi_graph_343_sft_e3_ck300 -> graph_her_nemi_sft_e3_final

her_nemi_graph_4096_sft_e3_ck524: relabel study: 4096 demos, batch 64, epoch 2 of 2 (eval on the box if time allowed; see CHECKPOINTS.md)

her_nemi_graph_4096_sft_e3_ck262: relabel study: 4096 demos (2731 reach + 1365 near, 3:4:3, per-prompt 2), batch 64, epoch 1 of 2 (not evaluated on the box)

her_nemi_graph_2048_sft_e3_ck260: relabel study: 2048 demos (1366 reach + 682 near, 3:4:3), batch 64, epoch 2 of 3; eval pass@8 10.8% (20.3/11.1/4.6), pass@1 1.90%

graph_nemi_rl_e3_best: pure RL stage 3 from graph_nemi_sft_e2, best eval pass@1 3.04% at step 280 (run paused at step 300)

her_nemi_graph_sft_e3_ck174: relabel study: main mix data, batch 64, epoch 2 of 3; eval pass@8 11.1% (20.9/13.4/2.8), pass@1 1.99%

her_nemi_graph_343_sft_e3_ck300: relabel study: 3:4:3 bands, 1024 reach + 512 near, batch 64, epoch 3 of 3; eval pass@8 12.0% (21.6/12.0/6.0), pass@1 2.17%

her_nemi_graph_sft_e3: relabel study: 1024 reach + 512 near (scene graph), from graph_nemi_sft_e2, 510 steps batch16; eval pass@8 12.0% (21.6/13.9/4.1), pass@1 1.92%

smoke_p2v_rl_final: final checkpoint (global_step_2 of 2 steps)

graph_nemi_sft_e2: SFT from scene_nemi_rl_e2_final on graph-stitched near-miss demos of the RL e2 rollouts; eval pass@8 11.1%

nemi_sft_e2

scene_nemi_rl_e2_best

scene_nemi_rl_e2_final

scene_nemi_sft_e1

scene_nemi_sft_e0

nemi_graph_mix_sft_e0

nemi_r3_sft_e0

nemi_all_sft_e0

nemi_sft_e0

her_cur_graph_prompt_sft_e0

her_cur_graph_sft_e0

her_cur_60_sft_e0

her_cur_sft_e0

her_unfiltered_sft_e0

her_unfiltered_epoch_0