anhtld commited on
Commit
009133e
·
verified ·
1 Parent(s): fb612cb

Auto-sync: 2026-06-30 04:59:14 (part 2)

Browse files
results/paper_table_status.json CHANGED
@@ -1467,6 +1467,44 @@
1467
  "best_config": null,
1468
  "gain_vs_h16_policy": 0.06260869565217392
1469
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1470
  {
1471
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
1472
  "label": "K4 compatible residual retrieval, near-miss challenger gate 0.03",
 
1467
  "best_config": null,
1468
  "gain_vs_h16_policy": 0.06260869565217392
1469
  },
1470
+ {
1471
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035",
1472
+ "label": "K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35",
1473
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_summary.json",
1474
+ "clean_deployment": "yes",
1475
+ "same_state_proposals": "no",
1476
+ "expert_proposal": "no",
1477
+ "story_role": "trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale",
1478
+ "fallback_success": null,
1479
+ "pending_job": "14955663/14955664",
1480
+ "path_exists": false,
1481
+ "status": "pending",
1482
+ "success": null,
1483
+ "std_success": null,
1484
+ "completed_seeds": null,
1485
+ "num_completed": null,
1486
+ "best_config": null,
1487
+ "gain_vs_h16_policy": null
1488
+ },
1489
+ {
1490
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040",
1491
+ "label": "K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35/0.40",
1492
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json",
1493
+ "clean_deployment": "yes",
1494
+ "same_state_proposals": "no",
1495
+ "expert_proposal": "no",
1496
+ "story_role": "trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales",
1497
+ "fallback_success": null,
1498
+ "pending_job": "14955665/14955666",
1499
+ "path_exists": false,
1500
+ "status": "pending",
1501
+ "success": null,
1502
+ "std_success": null,
1503
+ "completed_seeds": null,
1504
+ "num_completed": null,
1505
+ "best_config": null,
1506
+ "gain_vs_h16_policy": null
1507
+ },
1508
  {
1509
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
1510
  "label": "K4 compatible residual retrieval, near-miss challenger gate 0.03",
results/paper_table_status.md CHANGED
@@ -80,6 +80,8 @@ Baseline h=16 policy: 29.74%
80
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible residual retrieval, near-miss challenger gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
81
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible residual retrieval, near-miss challenger gate 0.005 | complete | 36.00% | +6.26 pp | yes | no | no | fine lower-margin calibration for the trace-motivated singleton near-miss challenger gate |
82
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible residual retrieval, near-miss challenger gate 0.015 | complete | 36.00% | +6.26 pp | yes | no | no | fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate |
 
 
83
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible residual retrieval, near-miss challenger gate 0.03 | complete | 35.94% | +6.20 pp | yes | no | no | upper-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
84
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 | complete | 35.94% | +6.20 pp | yes | no | no | trace-motivated challenger calibration test for whether wrong-gripper residuals carry conditional selector headroom |
85
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01 | complete | 35.59% | +5.86 pp | yes | no | no | revive high-precision singleton near-miss tangents without boosting toxic near-miss+no-op composites |
 
80
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible residual retrieval, near-miss challenger gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
81
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible residual retrieval, near-miss challenger gate 0.005 | complete | 36.00% | +6.26 pp | yes | no | no | fine lower-margin calibration for the trace-motivated singleton near-miss challenger gate |
82
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible residual retrieval, near-miss challenger gate 0.015 | complete | 36.00% | +6.26 pp | yes | no | no | fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate |
83
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35 | pending 14955663/14955664 | pending | pending | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale |
84
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35/0.40 | pending 14955665/14955666 | pending | pending | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales |
85
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible residual retrieval, near-miss challenger gate 0.03 | complete | 35.94% | +6.20 pp | yes | no | no | upper-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
86
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 | complete | 35.94% | +6.20 pp | yes | no | no | trace-motivated challenger calibration test for whether wrong-gripper residuals carry conditional selector headroom |
87
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01 | complete | 35.59% | +5.86 pp | yes | no | no | revive high-precision singleton near-miss tangents without boosting toxic near-miss+no-op composites |
scripts/build_paper_analysis.py CHANGED
@@ -425,6 +425,22 @@ METHODS = [
425
  "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p015_summary.json"
426
  ),
427
  ),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
428
  MethodSpec(
429
  key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
430
  label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.03",
 
425
  "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p015_summary.json"
426
  ),
427
  ),
428
+ MethodSpec(
429
+ key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035",
430
+ label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35",
431
+ summary_path=(
432
+ "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
433
+ "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_summary.json"
434
+ ),
435
+ ),
436
+ MethodSpec(
437
+ key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040",
438
+ label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35/0.40",
439
+ summary_path=(
440
+ "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
441
+ "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json"
442
+ ),
443
+ ),
444
  MethodSpec(
445
  key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
446
  label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.03",
scripts/build_paper_table_status.py CHANGED
@@ -785,6 +785,26 @@ SPECS = [
785
  story_role="fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate",
786
  pending_job="14955093/14955094",
787
  ),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
788
  ResultSpec(
789
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
790
  label="K4 compatible residual retrieval, near-miss challenger gate 0.03",
 
785
  story_role="fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate",
786
  pending_job="14955093/14955094",
787
  ),
788
+ ResultSpec(
789
+ key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035",
790
+ label="K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35",
791
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_summary.json",
792
+ clean_deployment="yes",
793
+ same_state_proposals="no",
794
+ expert_proposal="no",
795
+ story_role="trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale",
796
+ pending_job="14955663/14955664",
797
+ ),
798
+ ResultSpec(
799
+ key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040",
800
+ label="K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35/0.40",
801
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json",
802
+ clean_deployment="yes",
803
+ same_state_proposals="no",
804
+ expert_proposal="no",
805
+ story_role="trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales",
806
+ pending_job="14955665/14955666",
807
+ ),
808
  ResultSpec(
809
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
810
  label="K4 compatible residual retrieval, near-miss challenger gate 0.03",
scripts/eval_maniskill_policy_rollout.py CHANGED
@@ -240,6 +240,12 @@ def main(argv: list[str] | None = None) -> int:
240
  help="Optional comma-separated residual candidate types that may override the "
241
  "primary retrieval-residual selection in a second-stage challenger gate.",
242
  )
 
 
 
 
 
 
243
  parser.add_argument(
244
  "--retrieval-residual-challenger-margin",
245
  type=float,
@@ -290,6 +296,16 @@ def main(argv: list[str] | None = None) -> int:
290
  for item in args.retrieval_residual_challenger_types.split(",")
291
  if item.strip()
292
  )
 
 
 
 
 
 
 
 
 
 
293
  candidate_type_bonuses: dict[str, float] = {}
294
  try:
295
  for item in args.candidate_type_bonuses.split(","):
@@ -358,6 +374,7 @@ def main(argv: list[str] | None = None) -> int:
358
  retrieval_residual_direction=args.retrieval_residual_direction,
359
  retrieval_residual_reduce=args.retrieval_residual_reduce,
360
  retrieval_residual_challenger_types=retrieval_residual_challenger_types,
 
361
  retrieval_residual_challenger_margin=args.retrieval_residual_challenger_margin,
362
  lattice_exclude_types=lattice_exclude_types,
363
  candidate_type_bonuses=candidate_type_bonuses,
 
240
  help="Optional comma-separated residual candidate types that may override the "
241
  "primary retrieval-residual selection in a second-stage challenger gate.",
242
  )
243
+ parser.add_argument(
244
+ "--retrieval-residual-challenger-scales",
245
+ default="",
246
+ help="Optional comma-separated residual scales where challenger candidates may "
247
+ "override the primary retrieval-residual selection. Empty allows all scales.",
248
+ )
249
  parser.add_argument(
250
  "--retrieval-residual-challenger-margin",
251
  type=float,
 
296
  for item in args.retrieval_residual_challenger_types.split(",")
297
  if item.strip()
298
  )
299
+ try:
300
+ retrieval_residual_challenger_scales = tuple(
301
+ float(item.strip())
302
+ for item in args.retrieval_residual_challenger_scales.split(",")
303
+ if item.strip()
304
+ )
305
+ except ValueError as exc:
306
+ raise SystemExit(
307
+ "--retrieval-residual-challenger-scales must contain comma-separated floats"
308
+ ) from exc
309
  candidate_type_bonuses: dict[str, float] = {}
310
  try:
311
  for item in args.candidate_type_bonuses.split(","):
 
374
  retrieval_residual_direction=args.retrieval_residual_direction,
375
  retrieval_residual_reduce=args.retrieval_residual_reduce,
376
  retrieval_residual_challenger_types=retrieval_residual_challenger_types,
377
+ retrieval_residual_challenger_scales=retrieval_residual_challenger_scales,
378
  retrieval_residual_challenger_margin=args.retrieval_residual_challenger_margin,
379
  lattice_exclude_types=lattice_exclude_types,
380
  candidate_type_bonuses=candidate_type_bonuses,
scripts/slurm/eval_maniskill_policy_rollout.sbatch CHANGED
@@ -74,6 +74,10 @@ RETRIEVAL_RESIDUAL_CHALLENGER_TYPES="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES:-}"
74
  if [[ -n "${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON:-}" ]]; then
75
  RETRIEVAL_RESIDUAL_CHALLENGER_TYPES="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON//:/,}"
76
  fi
 
 
 
 
77
  RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN="${RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN:-0.0}"
78
  LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
79
  if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
@@ -154,6 +158,7 @@ apptainer exec --nv \
154
  --retrieval-residual-direction "$RETRIEVAL_RESIDUAL_DIRECTION" \
155
  --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
156
  --retrieval-residual-challenger-types "$RETRIEVAL_RESIDUAL_CHALLENGER_TYPES" \
 
157
  --retrieval-residual-challenger-margin "$RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN" \
158
  --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
159
  --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
 
74
  if [[ -n "${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON:-}" ]]; then
75
  RETRIEVAL_RESIDUAL_CHALLENGER_TYPES="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON//:/,}"
76
  fi
77
+ RETRIEVAL_RESIDUAL_CHALLENGER_SCALES="${RETRIEVAL_RESIDUAL_CHALLENGER_SCALES:-}"
78
+ if [[ -n "${RETRIEVAL_RESIDUAL_CHALLENGER_SCALES_COLON:-}" ]]; then
79
+ RETRIEVAL_RESIDUAL_CHALLENGER_SCALES="${RETRIEVAL_RESIDUAL_CHALLENGER_SCALES_COLON//:/,}"
80
+ fi
81
  RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN="${RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN:-0.0}"
82
  LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
83
  if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
 
158
  --retrieval-residual-direction "$RETRIEVAL_RESIDUAL_DIRECTION" \
159
  --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
160
  --retrieval-residual-challenger-types "$RETRIEVAL_RESIDUAL_CHALLENGER_TYPES" \
161
+ --retrieval-residual-challenger-scales "$RETRIEVAL_RESIDUAL_CHALLENGER_SCALES" \
162
  --retrieval-residual-challenger-margin "$RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN" \
163
  --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
164
  --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
scripts/slurm/summarize_h16_policy_ckpt.sbatch CHANGED
@@ -104,6 +104,9 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
104
  "retrieval_residual_challenger_types": data.get(
105
  "retrieval_residual_challenger_types", []
106
  ),
 
 
 
107
  "retrieval_residual_challenger_margin": data.get(
108
  "retrieval_residual_challenger_margin", 0.0
109
  ),
@@ -403,6 +406,16 @@ for row in rows:
403
  challenger=(
404
  " + challenger "
405
  + ",".join(row.get("retrieval_residual_challenger_types") or [])
 
 
 
 
 
 
 
 
 
 
406
  + f"@{row.get('retrieval_residual_challenger_margin') or 0.0:.2f}"
407
  if row.get("retrieval_residual_challenger_types")
408
  else ""
 
104
  "retrieval_residual_challenger_types": data.get(
105
  "retrieval_residual_challenger_types", []
106
  ),
107
+ "retrieval_residual_challenger_scales": data.get(
108
+ "retrieval_residual_challenger_scales", []
109
+ ),
110
  "retrieval_residual_challenger_margin": data.get(
111
  "retrieval_residual_challenger_margin", 0.0
112
  ),
 
406
  challenger=(
407
  " + challenger "
408
  + ",".join(row.get("retrieval_residual_challenger_types") or [])
409
+ + (
410
+ "["
411
+ + ",".join(
412
+ str(scale)
413
+ for scale in (row.get("retrieval_residual_challenger_scales") or [])
414
+ )
415
+ + "]"
416
+ if row.get("retrieval_residual_challenger_scales")
417
+ else ""
418
+ )
419
  + f"@{row.get('retrieval_residual_challenger_margin') or 0.0:.2f}"
420
  if row.get("retrieval_residual_challenger_types")
421
  else ""