kairos_press_button_full_train_16v3_reg_token_save_8pf_rms

kairos_press_button_full_train_16v3_reg_token_save_8pf_rms

Final checkpoints from a 3-phase training run on RMBench press_button (single task), testing a flat register-token memory bank for WAM ("V16.3-RT"). Trained from scratch on this task — no task-pretrained init. The video tower warm-starts from training/kairos/step-4500.safetensors.

Data

50 episodes, 26,352 trainable windows, LeRobot v2.1, mosaic camera, action space aloha_agilex_absolute_joint_14d (14-dim absolute joint). Topology: 1 node x 8 x A800-80GB.

Mechanism

A flat 64-token memory bank. Tokens are evicted independently rather than a whole frame-slot at a time, by

keep_i = region_rms_i * 0.5 ** (age_i / 8 chunks)

8 mosaic-aligned tokens are written per retained frame. The region grid is a 2x5 rectangular tiling — 12 regions of 10 patches on the 12x10 policy patch grid. Rectangular rather than square so that no region straddles the wrist-camera seam (patch column 5) or the wrist/head seam (patch row 4).

Phases

Each phase is a separate run directory with its own step counter. INIT_CKPT between phases is a warm start (begins at step 0 from those weights), not a resume — the 11K total steps below are 5000+5000+1000 across three directories, not one continuous counter.

Phase 1 is the control: it exists so that "memory helped" is measured against a backbone produced by this run on this task, not one inherited from another task.

Phase 2 is joint training and must be reported as one number. Its curve mixes "the towers got 5000 more steps" with "memory helped" — the two effects cannot be separated from it.

Phase 3's objective is not phase 1/2's objective. In phases 1-2 the logged loss is video + action (final step: loss=0.0020 loss_video=0.0018 loss_action=0.0002). In phase 3 (KAIROS_MEM_FINETUNE_ONLY=1) the logged loss equals loss_action exactly; video loss is reported but not optimized. Phase-3 loss is not comparable to phase-1/2 loss.

Phase 3 moved the loss very little: action loss ~0.0003 -> ~0.0002 over 1000 steps, and the reported (unoptimized) video loss stayed flat around 0.0016-0.0018. At the log's 4-decimal precision these differences are 1-2 quantisation units — the logs cannot resolve whether phase 3 improved anything.

What this does not establish

No held-out evaluation exists. EVAL_EVERY=0 in every phase, and this is enforced rather than merely unset: validate_kairos_memory_inline_eval_contract in training/little_wam/train.py raises for memory modes, because inline eval uses the stateless infer_action path, which does not reproduce the training-time memory cache. Every loss quoted above is training loss over the same 50 episodes, seen for roughly 833 epochs. These numbers speak to fit, not generalisation.

Not deployable as-is: the streaming counterpart for the flat bank is unwritten (StreamingSlotState refuses a flat-bank policy), so closed-loop inference from these weights is not yet supported.

token_half_life_chunks=8 and the gate threshold were set by argument, not measurement. Re-derive them from logs/surprise.rank*.jsonl with tools/measure_latent_surprise_rms.py before treating either as tuned.

Files

Weights (last checkpoint of each phase; no trainer state/optimizer files):

• p1_nomem_step_005000.pt
• p2_mem_step_005000.pt
• p3_memonly_step_001000.pt

Per-phase metadata (<phase>_latest.json, <phase>_train_args.json, <phase>_dataset_summary.json, <phase>_normalization_stats.json) and the configs/env scripts used to launch each phase are included alongside the weights.

Surprise logs

surprise_logs/ holds the per-chunk surprise score dumps, rank 0 only, one file per phase that produced them. Phase 1 produced none: its memory bank is off, so there is no surprise signal to dump. These are the inputs to tools/measure_latent_surprise_rms.py, which re-derives token_half_life_chunks and the gate threshold from a run (both were set by argument, not measurement, in this run — as noted above).

Provenance

Branch run/kairos-press-button-16v3-reg-token, commit 24d6e76b, ACP job pt-hu8vvvkq.