Upload RESULTS_MERGE_ACCURACY.md with huggingface_hub

RESULTS_MERGE_ACCURACY.md

@@ -1,6 +1,6 @@
 # Merging with alignment: does it improve DOWNSTREAM ACCURACY?
 
-_Generated 2026-08-27 00:52 UTC · training-free · code `/root/merge-accuracy` · merge operators, aligners and
+_Generated 2026-08-27 00:54 UTC · training-free · code `/root/merge-accuracy` · merge operators, aligners and
 quotient-distance diagnostics imported unmodified from `mergeschool.core` (`/root/mergeability`,
 treated as read-only)._
 
@@ -109,7 +109,7 @@ A merge that clears (a) but not (b) is not a usable model.
 |---|---|---|---|---|---|---|---|
 | IFEval prompt (strict) | 0.000 | 0.365 | 0.350 | 0.350 | **+0.000** | **no** | 0.540 |
 | IFEval instruction | 0.000 | 0.495 | 0.489 | 0.489 | **+0.000** | **no** | 0.640 |
-| Belebele ind_Latn | 0.250 | 0.643 | 0.660 | 0.660 | **+0.000** | yes | — |
+| Belebele ind_Latn | 0.250 | 0.643 | 0.660 | 0.660 | **+0.000** | yes | 0.690 |
 | Belebele eng_Latn | 0.250 | 0.747 | 0.783 | 0.783 | **+0.000** | yes | 0.847 |
 | ARC-easy | 0.250 | 0.728 | 0.726 | 0.726 | **+0.000** | **no** | 0.790 |
 
@@ -119,7 +119,7 @@ A merge that clears (a) but not (b) is not a usable model.
 |---|---|---|---|---|---|---|---|
 | IFEval prompt (strict) | 0.000 | 0.365 | 0.340 | 0.340 | **+0.000** | **no** | 0.540 |
 | IFEval instruction | 0.000 | 0.495 | 0.476 | 0.476 | **+0.000** | **no** | 0.640 |
-| Belebele ind_Latn | 0.250 | 0.643 | 0.620 | 0.620 | **+0.000** | **no** | — |
+| Belebele ind_Latn | 0.250 | 0.643 | 0.620 | 0.620 | **+0.000** | **no** | 0.690 |
 | Belebele eng_Latn | 0.250 | 0.747 | 0.747 | 0.747 | **+0.000** | **no** | 0.847 |
 | ARC-easy | 0.250 | 0.728 | 0.614 | 0.614 | **+0.000** | **no** | 0.790 |
 
@@ -159,7 +159,7 @@ A merge that clears (a) but not (b) is not a usable model.
 |---|---|---|---|---|---|---|---|
 | IFEval prompt (strict) | 0.000 | 0.140 | 0.475 | 0.475 | **+0.000** | yes | 0.540 |
 | IFEval instruction | 0.000 | 0.251 | 0.601 | 0.601 | **+0.000** | yes | 0.640 |
-| Belebele tha_Thai | 0.250 | 0.527 | 0.617 | 0.617 | **+0.000** | yes | — |
+| Belebele tha_Thai | 0.250 | 0.527 | 0.617 | 0.617 | **+0.000** | yes | 0.560 |
 | Belebele eng_Latn | 0.250 | 0.670 | 0.780 | 0.780 | **+0.000** | yes | 0.847 |
 | ARC-easy | 0.250 | 0.734 | 0.758 | 0.758 | **+0.000** | yes | 0.790 |
 
@@ -325,8 +325,8 @@ against an unpermuted reference of 0.375), so the mechanism is real and does rea
 | `Swallow_v02` | done | done | done | done |
 | `Typhoon2` | done | done | done | done |
 
-- reference `Llama-3.1-8B (base)`: belebele_jpn_Jpan 0.563, belebele_eng_Latn 0.743, arc_easy 0.728, ifeval_prompt 0.125, ifeval_inst 0.254
-- reference `Llama-3.1-8B-Instruct`: belebele_jpn_Jpan 0.720, belebele_eng_Latn 0.847, arc_easy 0.790, ifeval_prompt 0.540, ifeval_inst 0.640
+- reference `Llama-3.1-8B (base)`: belebele_ind_Latn 0.537, belebele_jpn_Jpan 0.563, belebele_tha_Thai 0.393, belebele_eng_Latn 0.743, arc_easy 0.728, ifeval_prompt 0.125, ifeval_inst 0.254
+- reference `Llama-3.1-8B-Instruct`: belebele_ind_Latn 0.690, belebele_jpn_Jpan 0.720, belebele_tha_Thai 0.560, belebele_eng_Latn 0.847, arc_easy 0.790, ifeval_prompt 0.540, ifeval_inst 0.640
 
 ### Supporting: a cross-group pair merged directly (not a chat vector)