suchirsalhan commited on
Commit
a402af7
·
verified ·
1 Parent(s): 8d7766e

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +30 -11
README.md CHANGED
@@ -355,15 +355,17 @@ Every table is also a CSV under `results/aim/tables/` and on the Hub at [`Mergea
355
 
356
  The direct test of AIM's claim. Needs no benchmark scores. `rho_W` is the weight-space shrink `‖ΔW_AIM‖/‖ΔW_merge‖`; `rho_A_*` is the same ratio measured on activations (4096 pile-val token positions, mean over layers); `aim_advantage = rho_A_uniform − rho_A_aim` is **positive when the activation-informed shrink preserves the base model's activations better than an activation-agnostic shrink of exactly the same weight-space size**. `omega_hat` and `col_r2_median` are the recovered closed form; `salience_recovery_r` is the correlation between the salience implied by the published checkpoint pair and the salience measured independently from the base model.
357
 
358
- | operator_label | combo | omega_hat | col_r2_median | scalar_r2_median | salience_recovery_r | rho_W | rel_dev_merge | rel_dev_uniform | rel_dev_aim | rho_A_uniform | rho_A_aim | aim_advantage | leverage_uniform | leverage_aim |
359
- |:-----------------|:----------|------------:|----------------:|-------------------:|----------------------:|--------:|----------------:|------------------:|:--------------|----------------:|:------------|:----------------|-------------------:|:---------------|
360
- | DARE Linear | Code-Math | 0.3971 | 0.9955 | 0.9949 | 0.9934 | 0.9239 | 0.209 | 0.2017 | | 0.9652 | | | 0.4579 | |
361
- | DARE TIES | Code-Math | 0.3961 | 0.9989 | 0.9984 | 0.9923 | 0.9256 | 0.2114 | 0.2053 | 0.1847 | 0.971 | 0.8739 | 0.09717 | 0.3893 | 1.695 |
362
- | Task Arithmetic | Code-Math | 0.3958 | 0.9988 | 0.9985 | 0.9921 | 0.9266 | 0.2079 | 0.2019 | 0.1835 | 0.9711 | 0.8828 | 0.08823 | 0.3946 | 1.597 |
363
- | TIES | Code-Math | 0.3939 | 0.9988 | 0.9984 | 0.99 | 0.9308 | 0.1834 | 0.1784 | 0.1624 | 0.9729 | 0.8855 | 0.08738 | 0.392 | 1.654 |
 
 
364
 
365
 
366
- **Paired test of the targeting advantage** over the magnitude-matched uniform control, across the 3 matched pairs: mean +0.09093, median +0.08823, 3/3 positive, Wilcoxon signed rank p = nan, sign test p = 0.25.
367
 
368
 
369
  ### T0 — The merge outcome, paired (published numbers, nothing re-run)
@@ -411,12 +413,29 @@ Per benchmark, paired across the 20 matched pairs:
411
 
412
  ### T2 — Does AIM's effect differ by merge operator?
413
 
414
- _Not yet computed (needs the property sweep)._
 
 
 
 
 
 
 
 
 
 
415
 
416
 
417
  ### T3 — Does it differ by which domains were merged?
418
 
419
- _Not yet computed (needs the property sweep)._
 
 
 
 
 
 
 
420
 
421
 
422
  ### T1 — Paired per-checkpoint table
@@ -431,7 +450,7 @@ One row per (operator, task combination, property): the with-AIM value, its base
431
  | operator_label | combo | ledger_status | weight_mechanism | activation_mechanism | property_panel_both_arms | published_benchmarks |
432
  |:-----------------|:----------------------------|:----------------|:-------------------|:-----------------------|:---------------------------|:-----------------------|
433
  | Task Arithmetic | Code-Math | not-run | True | True | False | True |
434
- | Task Arithmetic | Code-Instruction_Tuned | not-run | False | False | False | True |
435
  | Task Arithmetic | Math-Instruction_Tuned | not-run | False | False | False | True |
436
  | Task Arithmetic | Code-Math-Instruction_Tuned | not-run | False | False | False | True |
437
  | TIES | Code-Math | not-run | True | True | False | True |
@@ -446,7 +465,7 @@ One row per (operator, task combination, property): the with-AIM value, its base
446
  | DARE TIES | Code-Instruction_Tuned | not-run | False | False | False | True |
447
  | DARE TIES | Math-Instruction_Tuned | not-run | False | False | False | True |
448
  | DARE TIES | Code-Math-Instruction_Tuned | not-run | False | False | False | True |
449
- | WIDEN | Code-Math | not-run | False | False | False | True |
450
  | WIDEN | Code-Instruction_Tuned | not-run | False | False | False | True |
451
  | WIDEN | Math-Instruction_Tuned | not-run | False | False | False | True |
452
  | WIDEN | Code-Math-Instruction_Tuned | not-run | False | False | False | True |
 
355
 
356
  The direct test of AIM's claim. Needs no benchmark scores. `rho_W` is the weight-space shrink `‖ΔW_AIM‖/‖ΔW_merge‖`; `rho_A_*` is the same ratio measured on activations (4096 pile-val token positions, mean over layers); `aim_advantage = rho_A_uniform − rho_A_aim` is **positive when the activation-informed shrink preserves the base model's activations better than an activation-agnostic shrink of exactly the same weight-space size**. `omega_hat` and `col_r2_median` are the recovered closed form; `salience_recovery_r` is the correlation between the salience implied by the published checkpoint pair and the salience measured independently from the base model.
357
 
358
+ | operator_label | combo | omega_hat | col_r2_median | scalar_r2_median | salience_recovery_r | rho_W | rel_dev_merge | rel_dev_uniform | rel_dev_aim | rho_A_uniform | rho_A_aim | aim_advantage | leverage_uniform | leverage_aim |
359
+ |:-----------------|:-----------------------|------------:|----------------:|-------------------:|:----------------------|--------:|----------------:|:------------------|:--------------|:----------------|:------------|:----------------|:-------------------|:---------------|
360
+ | DARE Linear | Code-Math | 0.3971 | 0.9955 | 0.9949 | 0.9934 | 0.9239 | 0.209 | 0.2017 | 0.1832 | 0.9652 | 0.8767 | 0.08843 | 0.4579 | 1.62 |
361
+ | DARE TIES | Code-Math | 0.3961 | 0.9989 | 0.9984 | 0.9923 | 0.9256 | 0.2114 | 0.2053 | 0.1847 | 0.971 | 0.8739 | 0.09717 | 0.3893 | 1.695 |
362
+ | Task Arithmetic | Code-Instruction_Tuned | 0.3999 | 0.9995 | 0.9991 | | 0.9185 | 0.319 | 0.3071 | 0.2704 | 0.9626 | 0.8475 | 0.1151 | 0.4592 | 1.871 |
363
+ | Task Arithmetic | Code-Math | 0.3958 | 0.9988 | 0.9985 | 0.9921 | 0.9266 | 0.2079 | 0.2019 | 0.1835 | 0.9711 | 0.8828 | 0.08823 | 0.3946 | 1.597 |
364
+ | TIES | Code-Math | 0.3939 | 0.9988 | 0.9984 | 0.99 | 0.9308 | 0.1834 | 0.1784 | 0.1624 | 0.9729 | 0.8855 | 0.08738 | 0.392 | 1.654 |
365
+ | WIDEN | Code-Math | 0.3717 | 0.9899 | 0.9894 | — | 0.9352 | 0.15 | — | — | — | — | — | — | — |
366
 
367
 
368
+ **Paired test of the targeting advantage** over the magnitude-matched uniform control, across the 5 matched pairs: mean +0.09526, median +0.08843, 5/5 positive, Wilcoxon signed rank p = 0.0625, sign test p = 0.0625.
369
 
370
 
371
  ### T0 — The merge outcome, paired (published numbers, nothing re-run)
 
413
 
414
  ### T2 — Does AIM's effect differ by merge operator?
415
 
416
+ | operator | n_metrics | n_moved_q<0.05 | mean_|diff| (SD units) | largest movers |
417
+ |:-------------------|------------:|-----------------:|-------------------------:|:--------------------------------------------------------------------------------|
418
+ | ALL | 5 | 0 | 0.5915 | beh_js_to_base (-0.85), beh_kl_to_base (-0.82), act_deviation_from_base (-0.60) |
419
+ | DARETaskArithmetic | 5 | 0 | 0.5261 | beh_js_to_base (-0.89), beh_kl_to_base (-0.78), act_deviation_from_base (-0.53) |
420
+ | DARETies | 5 | 0 | 0.5311 | beh_js_to_base (-0.85), beh_kl_to_base (-0.78), act_deviation_from_base (-0.54) |
421
+ | TaskArithmetic | 5 | 0 | 0.7688 | beh_kl_to_base (-1.00), beh_js_to_base (-0.98), cka_to_base (+0.90) |
422
+ | Ties | 5 | 0 | 0.3762 | beh_js_to_base (-0.57), beh_kl_to_base (-0.56), act_deviation_from_base (-0.43) |
423
+ | WIDEN | 5 | 0 | 0.07494 | qmd_raw_from_base (-0.07), beh_js_to_base (+nan), beh_kl_to_base (+nan) |
424
+
425
+
426
+ Full table: `results/aim/tables/T2_per_operator.csv` (30 rows: every metric x every level, with the paired mean difference in SD units, the count of positive differences, Wilcoxon/sign p and BH q).
427
 
428
 
429
  ### T3 — Does it differ by which domains were merged?
430
 
431
+ | combo | n_metrics | n_moved_q<0.05 | mean_|diff| (SD units) | largest movers |
432
+ |:-----------------------|------------:|-----------------:|-------------------------:|:--------------------------------------------------------------------------------|
433
+ | ALL | 5 | 0 | 0.5915 | beh_js_to_base (-0.85), beh_kl_to_base (-0.82), act_deviation_from_base (-0.60) |
434
+ | Code-Instruction_Tuned | 5 | 0 | 1.029 | cka_to_base (+1.50), beh_kl_to_base (-1.22), beh_js_to_base (-1.09) |
435
+ | Code-Math | 5 | 0 | 0.4841 | beh_js_to_base (-0.79), beh_kl_to_base (-0.72), act_deviation_from_base (-0.50) |
436
+
437
+
438
+ Full table: `results/aim/tables/T3_per_combo.csv` (15 rows: every metric x every level, with the paired mean difference in SD units, the count of positive differences, Wilcoxon/sign p and BH q).
439
 
440
 
441
  ### T1 — Paired per-checkpoint table
 
450
  | operator_label | combo | ledger_status | weight_mechanism | activation_mechanism | property_panel_both_arms | published_benchmarks |
451
  |:-----------------|:----------------------------|:----------------|:-------------------|:-----------------------|:---------------------------|:-----------------------|
452
  | Task Arithmetic | Code-Math | not-run | True | True | False | True |
453
+ | Task Arithmetic | Code-Instruction_Tuned | not-run | True | True | False | True |
454
  | Task Arithmetic | Math-Instruction_Tuned | not-run | False | False | False | True |
455
  | Task Arithmetic | Code-Math-Instruction_Tuned | not-run | False | False | False | True |
456
  | TIES | Code-Math | not-run | True | True | False | True |
 
465
  | DARE TIES | Code-Instruction_Tuned | not-run | False | False | False | True |
466
  | DARE TIES | Math-Instruction_Tuned | not-run | False | False | False | True |
467
  | DARE TIES | Code-Math-Instruction_Tuned | not-run | False | False | False | True |
468
+ | WIDEN | Code-Math | not-run | True | True | False | True |
469
  | WIDEN | Code-Instruction_Tuned | not-run | False | False | False | True |
470
  | WIDEN | Math-Instruction_Tuned | not-run | False | False | False | True |
471
  | WIDEN | Code-Math-Instruction_Tuned | not-run | False | False | False | True |