SparseTrajTokens β€” ablation and search checkpoints

Final checkpoints of runs that are not the released model (see SparseTrajTokens for the four released ones). Kept for reproducibility of the ablations and of the schedule search; every run has its meta.json (all hyper-parameters), train_log.jsonl and, where applicable, val_curve.jsonl / PREREG.md. Optimizer states and intermediate steps are not included.

Group Runs What
runs_sft110k/ pB_rgbsw_3000_s0, pB_rgbsw_6000_s0 frames-only budget sweep (the 9,002-step arm is the released rgb_only_arm)
pB_f_tok_6k_s0, pB_f_s2_1500_s0, pB_f_s2_3000_s0 the previous (cosine) stage-2 / stage-3 recipe
pB_d2_no1a_s{0,1}, pB_d2_with1a_s{0,1} value of the stage-1 warm-up, two seeds each
pB_rgb_S_e8_s0, pB_traj_only_S_e8_s0, 1a_concat_tsec_v4td earlier structured-target arms and the earlier stage-1 projector
runs_t2/ 35 stage-1 projector runs tokenization-structure ablations (fusion / time axis / camera), see PAPER_TOKENIZATION.md in the archive
runs_ablation_0824/ 14 runs August tokenization ablations

Evaluation outputs of all these runs are in TrajQA-B/results/ and in the workspace archive (sparsetraj-archive).

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Chenfy0407/SparseTrajTokens-ablations

Adapter
(257)
this model