Upload folder using huggingface_hub
Browse files- logs/plans/MI_TUNE_PLAN.md +67 -0
- logs/plans/OVERNIGHT2_PLAN.md +17 -0
- logs/plans/OVERNIGHT_PLAN.md +37 -0
logs/plans/MI_TUNE_PLAN.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Overnight MI-weight auto-tune (2026-09-12, user asleep)
|
| 2 |
+
Recipe: sparse binary (gate-L0 2e-3), hot (lr 3e-4 const, tau 0.3), balanced 869-task pool, tpi 32 / K 4, batch 128, bf16+compile, 60k steps.
|
| 3 |
+
MI term = w * sum_{i<j} MI(z_i; z_j) (batch estimate; ~2e3 at init, ~2e4 @1k, ~4.5e5 for a trained sparse code).
|
| 4 |
+
Round 1: w in {0 (control, GPU2), 3e-8, 3e-7 (GPU0), 3e-6, 3e-5 (GPU1)}. Stage-3 chain (3 head seeds + search full-0.5pt) after each.
|
| 5 |
+
Rules:
|
| 6 |
+
- 20k eval: bitmatch MEAN < 90 or bacc < 80 -> kill (too strong); replace with geometric midpoint between it and the largest surviving w.
|
| 7 |
+
- Score (Stage-3): primary = dense-head OOD mean over 3 seeds; secondary = search OOD (and |S|); constraint = all six bitmatch >= 95, presence reported.
|
| 8 |
+
- Round 2: bracket the best w at x/sqrt(10) and x*sqrt(10) (untried values only) + seed 2 of the best; Stage-3 chains; final table.
|
| 9 |
+
Markers in runs/lane_scale.log; per-run logs runs/vqa_bal_mi<w>.log, Stage-3 runs/stage3_vqa_bal_mi<w>.log, search runs/search_vqa_bal_mi<w>_d005.log.
|
| 10 |
+
|
| 11 |
+
## Round 1 results (60k, seed 1; head = dense-head OOD over 3 seeds; search = full-0.5pt)
|
| 12 |
+
| w | bitmatch | presence | head OOD (3 seeds) mean | drop | search |S| / OOD |
|
| 13 |
+
| 0 | 97.4 (shape 94.4) | 94.8 | 84.8 87.2 84.1 -> 85.4 | +13.4 | 21 / 96.9 |
|
| 14 |
+
| 3e-8 | 99.4 | 96.1 | 81.1 76.4 85.9 -> 81.1 | +18.2 | 150 / 63.1 |
|
| 15 |
+
| 3e-7 | 98.8 | 89.8 | 92.9 97.5 95.9 -> 95.4 | +3.9 | 10 / 98.4 <- best |
|
| 16 |
+
| 3e-6 | 97.2 (shape 91.3) | 95.7 | 65.7 76.0 68.7 -> 70.1 | +29.0 | 5 / 60.1 |
|
| 17 |
+
| 3e-5 | 99.5 | 98.0 | 67.0 74.7 65.2 -> 69.0 | +30.4 | 5 / 97.6 |
|
| 18 |
+
Note: the w=0 control at tpi32/K4/batch128 is far below the old tpi8/K16/batch192 sparse baseline (94.4 / 10 cols 97.8) -> recipe change costs purity (shape bitmatch 94.4 too); 3e-8 < 0 shows single-seed noise is large.
|
| 19 |
+
Round 2 (launched 14:50 UTC): 1e-7 s1, 1e-6 s1 (GPU0); 3e-7 s2, 0 s2 (GPU1); 3e-5 s2, 3e-7 s3 (GPU3). detr-1024 + tpi32k4-control Stage-3 on GPU2.
|
| 20 |
+
MI of the round-1 codes measured post hoc on the same 11264 slots (pairwise_mi on sigmoid probs; also mean p and fraction of saturated bits |p-0.5|>0.45):
|
| 21 |
+
| w | MI sum (nats) | per-pair | mean p | saturated |
|
| 22 |
+
| 0 | 622926 | 7.4e-2 | 0.500 | 89.6% |
|
| 23 |
+
| 3e-8 | 2006 | 2.4e-4 | 0.197 | 88.9% |
|
| 24 |
+
| 3e-7 | 242 | 2.9e-5 | 0.226 | 78.7% |
|
| 25 |
+
| 3e-6 | 111 | 1.3e-5 | 0.881 | 4.7% | <- probs pinned near 0.88 (constant across slots = MI 0), barely binary
|
| 26 |
+
| 3e-5 | 27 | 3.2e-6 | 0.012 | 99.1% | <- ~1% of bits active: MI minimized via near-zero marginals (act-sparsity regime)
|
| 27 |
+
refs: old sparse (tpi8/K16) 454151 / 5.4e-2 / p 0.501 / 92.6%; dense 174582 / 2.1e-2.
|
| 28 |
+
Gate usage (869 tasks; open = p>0.5): columns used by >=1 task / open per task (median [min-max]) / tasks per used column (median)
|
| 29 |
+
| 0: 122 / 9 [3-19] / 43 | 3e-8: 119 / 10 [6-21] / 40 | 3e-7: 91 / 18 [13-29] / 54 | 3e-6: 106 / 9 [4-21] / 43 | 3e-5: 101 / 8 [2-20] / 53 |
|
| 30 |
+
refs: old sparse 101 / 7 / 44; dense 2881 / 791 / 148; cont 67 / 8 / 54; cont-no-pen 2202 / 474 / 94; detr-512 89 / 11 / 54
|
| 31 |
+
|
| 32 |
+
## Round 2 results (head OOD = 3-seed mean; search = full-0.5pt)
|
| 33 |
+
| arm | bitmatch | presence | head OOD | drop | |S| / OOD |
|
| 34 |
+
| 3e-7 s1 (r1) | 98.8 | 89.8 | 95.4 | +3.9 | 10 / 98.4 |
|
| 35 |
+
| 3e-7 s2 | 99.2 | 90.6 | 83.1 | +16.4 | 48 / 98.2 |
|
| 36 |
+
| 3e-7 s3 | 99.6 | 97.3 | 88.0 | +11.6 | 196 / 98.4 |
|
| 37 |
+
| 0 s1 (r1) | 97.4 | 94.8 | 85.4 | +13.4 | 21 / 96.9 |
|
| 38 |
+
| 0 s2 | 99.4 | 97.1 | 92.7 | +6.9 | 280 / 91.6 |
|
| 39 |
+
| 1e-7 s1 | 98.5 | 93.7 | 73.5 | +26.0 | 8 / 95.9 |
|
| 40 |
+
| 1e-6 s1 | 99.2 | 95.6 | 70.6 | +28.8 | 5 / 99.2 |
|
| 41 |
+
| 3e-5 s1 (r1) | 99.5 | 98.0 | 69.0 | +30.4 | 5 / 97.6 |
|
| 42 |
+
| 3e-5 s2 | 98.7 | 90.8 | 80.6 | +18.6 | 5 / 98.1 |
|
| 43 |
+
| w=0, batch 192, 120k, s2 (tpi32k4 control) | 98.5 | - | 94.7 | +5.0 | 26 / 98.8 |
|
| 44 |
+
| detr-1024 lr1e-4 (tpi8/K16, b128, 120k, s2) | 99.5 | - | 91.1 | +7.9 | 358 stall / 64.1 |
|
| 45 |
+
Verdict so far: dense-head OOD is dominated by seed noise at 60k/b128 (3e-7: 95.4/83.1/88.0; 0: 85.4/92.7) -> MI 3e-7 is NOT a robust head win.
|
| 46 |
+
Robust MI effect at w >= 1e-6: 5-column supports that transfer (99.2/98.1/97.6; one failure 60.1 at 3e-6 s1) + selection-free head collapse (70-80).
|
| 47 |
+
Round 3 (launched 22:05 UTC): 1e-6 s2/s3, 3e-6 s2, 3e-5 s3 (60k/b128); 3e-7 and 1e-6 at b192/120k s2 (vs tpi32k4 control).
|
| 48 |
+
|
| 49 |
+
## Round 3 (60k/b128 replicates): 1e-6 s2 60.0 head / 5 cols @98.3; 1e-6 s3 81.8 / 6 @98.7; 3e-6 s2 79.5 / 5 @98.8; 3e-5 s3 60.5 / 4 @99.1.
|
| 50 |
+
=> w >= 1e-6 regime robust (8/9 arms): 4-6 col support @98-99, head 60-82. Round 4 (launched 05:45 UTC): 3e-4, 1e-3, 3e-3 (seed 1).
|
| 51 |
+
|
| 52 |
+
## b192/120k seed 2: 3e-7 -> head 87.2 (+12.3), search 99 cols @98.4, bitmatch 97.7 (shape 89.5); 1e-6 -> head 80.1 (+19.4), search 7 cols @98.1, bitmatch 99.5, presence 98.7.
|
| 53 |
+
## Round 4 (3e-4, 1e-3, 3e-3): all collapse — bitmatch 80-84 (majority floor), presence 55-68, bacc 70-91; MI term wins over CE (act pinned 0.88 / 0.51 / 0.006). Upper bound of usable w ~ 3e-5..1e-4.
|
| 54 |
+
## No-gate baseline (gate==1, no penalties) @28k/38k: binary 96.3 (posX/Y 91.8), continuous 95.3 (size 87.4) — disentangles without any gate; slower on position/size.
|
| 55 |
+
|
| 56 |
+
## Support search with 10k in-rule scenes (val fixed 5050, bar full-0.5pt) vs 45k
|
| 57 |
+
| code | 45k |S| @OOD | 10k |S| @OOD |
|
| 58 |
+
| sparse (gate-L0) | 10 @97.8 | 3 @98.2 |
|
| 59 |
+
| dense binary | 21 @98.4 | 16 @98.8 |
|
| 60 |
+
| continuous (gate-L0) | 45 @83.2 | 64 @86.4 (stall) |
|
| 61 |
+
| tpi32k4 control (b192) | 26 @98.8 | 5 @98.8 |
|
| 62 |
+
| MI 0 (60k s1) | 21 @96.9 | 9 @63.7 <- spurious |
|
| 63 |
+
| MI 3e-7 | 10 @98.4 | 5 @98.4 |
|
| 64 |
+
| MI 1e-6 s1 | 5 @99.2 | 12 @98.6 |
|
| 65 |
+
| MI 3e-5 | 5 @97.6 | 4 @98.9 |
|
| 66 |
+
=> 10k is enough for the oracle (full-code fidelity unchanged, 98.5 vs 98.3); supports get SMALLER for binary codes (noisier oracle
|
| 67 |
+
-> looser removal test -> more peeling) and one spurious pass (MI 0). The sparse code's 3-column support @98.2 = literals only.
|
logs/plans/OVERNIGHT2_PLAN.md
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Overnight 2 (2026-09-14 → 15): find where redundant / correlated features actually hurt (sparse+MI should win)
|
| 2 |
+
|
| 3 |
+
Codes (ladder): sparse=vqa_balanced · sparseMI=vqa_bal_mi1e-6_s1 · dense=vqa_balanced_dense · dense60k=vqa_bal_dense60k_s2 ·
|
| 4 |
+
cont=vqa_bal_cont (CONT) · MIonly=vqa_bal_miNoL0_1e-6_s1 · lin_gate=vqa_bal_lin_sgate_none_s1 (CONT) · sparse_b192=vqa_bal_s2_tpi32k4
|
| 5 |
+
|
| 6 |
+
Hypotheses (each is a lane; all use the single-task probe unless stated):
|
| 7 |
+
A. Data efficiency under the spurious rule (cube→gray, "small metal cubes"): NTR ∈ {500, 2000} in-rule scenes for search + head.
|
| 8 |
+
Prediction: dense/loose-copy codes need many scenes to reject aliases; sparse+MI's 5 clean columns hold at 500.
|
| 9 |
+
B. Composite-column adversary: dataset cube_region (cube ⇔ x7∨y0) — dense60k has a single column ≈ x7∨y0 (census). Task "count cubes".
|
| 10 |
+
Prediction: dense60k's search/head can route to the composite column (same sparsity as the cube literal) → OOD failure; codes
|
| 11 |
+
without that column are fine. cube_red (cube → red), task "count cubes": alias needs 2 columns → search fine, head may mix in red.
|
| 12 |
+
C. Multi-rule datasets multi3 / multi2: v2 OOD multi-task ladder (targeted alias tasks for each attr→attr rule) + single-task probes
|
| 13 |
+
aliased by two rules at once (multi3: "count small cubes" — in-rule = small∧cube = small∧cube∧gray∧metal).
|
| 14 |
+
D. 13-dataset sweep of the v2 OOD multi-task (5k steps) for sparse, sparseMI, dense: mean disagreement accuracy per code.
|
| 15 |
+
E. Tight fidelity bar (full−0.2pt) search across codes on cube→gray at NTR=10k: does dense's duplicate web stall while sparse+MI stays at 5?
|
| 16 |
+
|
| 17 |
+
Outputs: runs/ovn2/*.log|json ; markers in runs/lane_ovn2.log ; summary via runs/ovn2_summary.py
|
logs/plans/OVERNIGHT_PLAN.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Overnight autonomous iteration (2026-09-10 night) — goals & state
|
| 2 |
+
## Goals (user-confirmed)
|
| 3 |
+
1. VQA fullpool: ALL SIX attribute bitmatches >= 95 (incl posX/posY, G=8). Allowed: value-balancing, literal-only families.
|
| 4 |
+
2. Robotic: MULTI-TASK instruction-conditioned closed-loop success in-distribution. pos-bypass config counts; strict all-bits = stretch (report both).
|
| 5 |
+
## Current best
|
| 6 |
+
- VQA: hotB_90k = 94.4 MEAN (color 96.8 / shape 89.4 [peaked 97.8@50k then drifted!] / size 99.7 / mat 99.6 / posX 91.0 / posY 89.8). Drift trades attrs at const-lr — may need cool-down tail to HOLD an all->=95 allocation.
|
| 7 |
+
- Robotic: bc_ap_v3.pt (1.84M steps, F1024, recon5, pos-bypass, 150k) = MULTI-TASK CLOSED-LOOP 8/15/12/12 per 20 (~59% avg) — GOAL ESSENTIALLY MET in pos-bypass config. Code: shape 99.3/color 99.8/size 99.8, pos probes 80/85.
|
| 8 |
+
## Lanes
|
| 9 |
+
- GPU2 vqa_litonly: literal-only families (zh_count,zh_exist,or_count,or_exist,cmp_int), hot recipe, 120k, hourly snapshots (runs/lane_vqaA.log marker VQA-LITONLY DONE).
|
| 10 |
+
- GPU3 lane B: hotB plateau search (marker PLATEAU SEARCH DONE) then vqa_balanced: fullpool + --balance-values, hot, 120k (marker VQA-BALANCED DONE).
|
| 11 |
+
- GPU0 bc_ap_v3strict: strict all-bits (no pos-bypass), big data, 150k + closed-loop (marker V3STRICT DONE).
|
| 12 |
+
- GPU1 bc_ap_v3n: act-noise 0.08 demos (DAgger-lite, expert actions recorded) -> pos-bypass 150k + closed-loop (marker V3NOISE DONE).
|
| 13 |
+
## Iteration protocol
|
| 14 |
+
- On any VQA eval with all six >=95: cp ckpt to runs/GOAL_<name>.pt IMMEDIATELY (drift!), then continue + optionally 1e-4-const cool-down resume to hold allocation.
|
| 15 |
+
- If litonly/balanced lag hotB trajectory at 40k: kill, try combo (litonly+balanced) or fullpool+cool-down-from-hotB_90k (resume hotB at 1e-4 const).
|
| 16 |
+
- Robotic: compare v3n vs v3 (59%); if noise >= +10pts, next = noise+300k or noise+more tasks. v3strict = report only.
|
| 17 |
+
- Ops rules: kill by PID + verify nvidia-smi release; log-marker chains only; eval scripts read flags/n_tasks from ckpt.
|
| 18 |
+
|
| 19 |
+
## Round 2 (after first completions)
|
| 20 |
+
- v3strict RESULT: all-bits closed-loop 1/6/4/8 per 20 (~24% avg) at bitmatch 94.9 — FIRST nonzero strict result; tracks code quality.
|
| 21 |
+
- v3noise RESULT: 5/3/7/2 (~21%) — act-noise 0.08 HURT (vs v3 59%); noise lever dead at this magnitude.
|
| 22 |
+
- Lanes now: GPU0 v4 = pos-bypass 300k (champion extension, target >59%); GPU1 v4strict = all-bits F=2048 300k (stretch scale-up); GPU2 litonly r3 (preflight-guarded after 2 OOM collisions — contp zombie then dual-instance race); GPU3 plateau search -> vqa_balanced.
|
| 23 |
+
- Ops: kill CHAIN WRAPPERS by pid too; preflight GPU-free check before any launch; single monitor at a time.
|
| 24 |
+
|
| 25 |
+
## GOAL 1 MET (VQA all six >= 95)
|
| 26 |
+
- vqa_balanced (--balance-values full pool, 869 tasks ~50/value, hot recipe): ALL SIX >= 95 by step 20k;
|
| 27 |
+
@30k MEAN 98.6 (color 99.9 shape 99.2 size 98.9 material 99.7 posX 98.1 posY 95.8). Ckpt: runs/GOAL_vqa_balanced_allsix.pt (@30k snapshot).
|
| 28 |
+
Value BALANCING was the bottleneck (positions/color starved in natural pool). Run continues to 120k; hourly snapshots.
|
| 29 |
+
- litonly (96.5, pos 92.7/93.6) = second place; keep to completion as ablation (literal-only helps but balancing is the key).
|
| 30 |
+
## Round 3
|
| 31 |
+
- v4 full-14-task closed-loop eval running (GPU0) -> final multi-task table.
|
| 32 |
+
- v5strict F=4096 relaunched after bitmatch-eval OOM fix (cap 20k val frames).
|
| 33 |
+
- v4 (pos-bypass 300k) = 60% avg / 80% on position task; strict curve 24% -> 32.5% -> (F4096 pending).
|
| 34 |
+
|
| 35 |
+
## Definitive numbers
|
| 36 |
+
- ROBOTIC MULTI-TASK (v4, pos-bypass, all 14 tasks x 20 eps): 194/280 = 69.3% avg (min 50% cube, max 90% posY:2). GOAL MET WITH FULL TABLE.
|
| 37 |
+
- vqa_balanced @40k: MEAN 98.9 (posX 98.0, posY 96.4) — goal ckpt refreshed.
|