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 2026-08-26 21:46 UTC · training-free · code `/root/merge-accuracy` · merge operators, aligners and
+_Generated 2026-08-26 22:14 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)._
 
@@ -27,6 +27,11 @@ which forks need that.
 
 ## Headline
 
+**1 real community fork(s), 0 ground-truth control(s).**
+
+- On **1 of 1** real community CPT forks the fitted alignment map is the **identity** (coordinate share exactly 0; every per-layer MLP and attention-head permutation comes back as the identity). Continued pretraining by a third party did **not** move these models out of the base model's frame, so the chat vector is already expressed in the right basis and aligning it is a no-op. The measured accuracy difference is **exactly zero on every benchmark** — the aligned and naive merges are bit-identical models.
+- The chat-vector recipe itself **works** on 1 of 1 of these forks: it lifts instruction following well above the fork it started from, i.e. the merged model beats its own parent — the bar that matters.
+- The diagnostic's registered prediction was correct on **1/1** real forks.
 
 ## Substrate
 
@@ -34,6 +39,7 @@ which forks need that.
 |---|---|---|
 | base | `meta-llama/Llama-3.1-8B` | Meta |
 | instruct | `meta-llama/Llama-3.1-8B-Instruct` | Meta — the chat vector is Instruct − Base |
+| fork | `tokyotech-llm/Llama-3.1-Swallow-8B-v0.1` | community CPT fork, target `jpn_Jpan` |
 
 Every fork is shape-identical to the base (vocab 128256, hidden 4096, 32 layers, 32 heads / 8 KV
 heads), so the chat vector is added to **all 291 tensors**, embeddings included. Shared ancestry
@@ -63,25 +69,61 @@ decision variable. The factor columns show which parts of `g` survived the accep
 `MLP perm = id` says whether the accepted per-layer permutation was in fact the identity (a factor
 can be accepted and still be the identity, since equality passes the `<=` test).
 
-| model | coord. share | MLP perm = id | resid. factor | head factor | CKA vs base | rel. drift | weight cos | **PREDICTION** | fit cost |
+| model | coord. share | g = identity? | resid. factor | head perm = id | CKA vs base | rel. drift | weight cos | **PREDICTION** | fit cost |
 |---|---|---|---|---|---|---|---|---|---|
+| `Swallow` | **0.0000** | yes | kept | yes | 0.997 | 0.1784 | 0.9870 | do not align | 2104s |
 
 ## 2. Accuracy — fork alone / naive chat vector / aligned chat vector
 
 Bars to clear: **(a)** aligned beats naive; **(b)** the merged model beats the fork it came from.
 A merge that clears (a) but not (b) is not a usable model.
 
+### `Swallow`  ·  λ=1.0  ·  coord. share 0.0000  ·  prediction: do not align
+
+| metric | chance | fork alone | naive | aligned | Δ align | beats fork? | Instruct ref |
+|---|---|---|---|---|---|---|---|
+| IFEval prompt (strict) | 0.000 | 0.175 | 0.375 | 0.375 | **+0.000** | yes | 0.540 |
+| IFEval instruction | 0.000 | 0.315 | 0.514 | 0.514 | **+0.000** | yes | 0.640 |
+| Belebele jpn_Jpan | 0.250 | 0.600 | 0.680 | 0.680 | **+0.000** | yes | 0.720 |
+| Belebele eng_Latn | 0.250 | 0.680 | 0.790 | 0.790 | **+0.000** | yes | 0.847 |
+| ARC-easy | 0.250 | 0.746 | 0.772 | 0.772 | **+0.000** | yes | 0.790 |
+
+## 3. Selection experiment
+
+| strategy | alignment compute | mean IFEval prompt acc | pairs aligned | compute saved |
+|---|---|---|---|---|
+| merge naive (never align) | 0s | **0.3750** | 0/1 | 100% |
+| align everything | 2104s | **0.3750** | 1/1 | 0% |
+| diagnose -> align if coord_share >= 0.01 | 0s | **0.3750** | 0/1 | 100% |
+
 ## 4. Coverage
 
 | model / cell | diagnostic | fork alone | naive | aligned |
 |---|---|---|---|---|
-| `Swallow` | — | done | — | — |
+| `Swallow` | done | done | done | done |
 | `Swallow perm 0.0625` | — | done | — | — |
+| `Swallow perm 0.5` | — | done | — | — |
+| `Swallow perm 1.0` | — | done | — | — |
 | `Swallow_v02` | — | 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
 
+### Supporting: a cross-group pair merged directly (not a chat vector)
+
+`EleutherAI/pythia-1.4b` (step143000) x `SJTU-CL/Zh-Pythia-1.4B` — same architecture, different
+group, different tokenizer, no shared ancestor (weight cosine ~0). Body-only weight average,
+naive vs permutation/orthogonal aligned, scored on SciQ / PIQA / ARC-easy / LAMBADA.
+
+| arm | alpha | mean acc | parent A | parent B |
+|---|---|---|---|---|
+| naive | 0.25 | 0.2490 | 0.6665 | 0.3995 |
+| aligned | 0.25 | 0.2505 | 0.6665 | 0.3995 |
+| naive | 0.5 | 0.2480 | 0.6665 | 0.3995 |
+| aligned | 0.5 | 0.2480 | 0.6665 | 0.3995 |
+| naive | 0.75 | 0.2650 | 0.6665 | 0.3995 |
+| aligned | 0.75 | 0.2500 | 0.6665 | 0.3995 |
+
 ## 4. Method notes and a bug found in the shared library
 
 The alignment map `g` is fitted from **(fork, base)** — the map carrying the base model's