| --- |
| license: apache-2.0 |
| tags: |
| - model-merging |
| - model-alignment |
| - chat-vector |
| - llama-3.1 |
| --- |
| |
| # merge-accuracy — does aligning a merge improve DOWNSTREAM ACCURACY? |
|
|
| The mergeability line of work measures merge obstruction in **nats/token**. This dataset supplies |
| the missing axis: **task accuracy**, on real released models, for the merge recipe practitioners |
| actually run — the **chat-vector recipe** |
|
|
| ``` |
| theta_new = theta_fork + lambda * ( theta_instruct - theta_base ) |
| ``` |
|
|
| with `meta-llama/Llama-3.1-8B`, its official Instruct release, and **three community |
| continued-pretrained language forks from three independent groups** (Swallow / Japanese, |
| Typhoon2 / Thai, SEA-LION / Indonesian), plus **ground-truth permutation controls** and one |
| **cross-group direct-merge pair** (`pythia-1.4b` x `Zh-Pythia-1.4B`). |
|
|
| ## Headline |
|
|
| > **Alignment did not improve downstream accuracy on a single real model — because on every |
| > real model there was nothing to align.** Every community fork of `Llama-3.1-8B` we tested is |
| > still exactly in the base model's coordinate frame, so the aligned and |
| > naive chat vectors are bit-identical models and the accuracy difference is 0.000 on every |
| > benchmark. The diagnostic said so **before** any merge was built, and a 44-second screen |
| > reproduces that call 37x cheaper than fitting the map. When the frame really has drifted — |
| > a real fork acted on by a random element of its own symmetry group — the naive chat vector |
| > collapses (IFEval 0.175 -> 0.110, *below* the fork it started from) and alignment restores |
| > it to 0.355 against an unpermuted reference of 0.375. **The mechanism is real and does reach |
| > accuracy; the ecosystem condition that would make it pay off did not occur in any released |
| > model we examined.** |
|
|
| Population: **4 real community forks** measured end-to-end on accuracy (3 groups, 3 target languages), **3 ground-truth controls**, one cross-group direct-merge pair, and a **10-model ecosystem screen**. Across all **13 released `Llama-3.1-8B` derivatives** examined — language forks, domain continued pretraining, instruct post-training, a safety model — **not one** had left the base model's coordinate frame. |
|
|
| - On **4 of 4** 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 3 of 4 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 **4/4** real forks. |
| - On the ground-truth control (a real fork acted on by a random element of the model's own symmetry group — functionally identical, differently parameterised), the diagnostic fires (coordinate share **0.856**), the naive chat vector scores IFEval **0.110**, and aligning it first recovers **0.355** (Δ **+0.245**). |
|
|
| ## Contents |
|
|
| | path | what | |
| |---|---| |
| | `RESULTS_MERGE_ACCURACY.md` | the full write-up: substrate, chance levels, every table, limitations | |
| | `figs/headline_diagnostic_vs_gain.png` | **headline** — pre-merge coordinate share vs realised accuracy gain from aligning | |
| | `figs/dose_response.png` | the diagnostic tracks real frame drift; alignment recovers what drift destroys | |
| | `figs/scatter_naive_vs_aligned.png` | naive vs aligned accuracy with the `y = x` diagonal and the fork-alone baseline | |
| | `figs/selection_experiment.png` | naive / align-everything / cheap-screen-then-align | |
| | `results/chatvec_summary.csv` | per model per lambda: fork alone, naive, aligned, delta, per benchmark | |
| | `results/all_model_accuracies.csv` | every model evaluated, every benchmark | |
| | `results/diagnostics.csv` | the pre-merge diagnostic for every model | |
| | `results/cheap_screen_vs_full.csv` | the 44-second screen against the 27-minute fit | |
| | `results/selection_experiment.csv` | the three-strategy comparison with measured compute | |
| | `results/ecosystem_screen.json` | the 44-second frame screen over 10 more released Llama-3.1-8B derivatives | |
| | `figs/ecosystem_drift_vs_frame.png` | parameter drift vs frame drift across the ecosystem | |
| | `results/crossgroup_pair.csv` | the pythia x Zh-Pythia direct merge, all mixing weights + TIES | |
| | `results/validation.json` | exactness of the symmetry action and of the fitted alignment map | |
| | `results/chatvec.jsonl`, `results/ledger.jsonl` | raw resumable ledgers | |
| | `code/` | everything needed to reproduce | |
|
|
| ## Benchmarks and chance levels |
|
|
| Belebele (target language and English) — chance **0.250**; ARC-easy — chance **0.250**; |
| IFEval strict prompt-level and instruction-level — chance **~0**. |
| `lm-evaluation-harness` was not available, so scorers are implemented directly following the |
| harness / reference task definitions (`code/tasks.py`, `code/ifeval.py`). Sanity check: the |
| loglikelihood harness scores `EleutherAI/pythia-1.4b` at SciQ **0.846** against a published |
| **0.865**. |
|
|
| ## A bug worth propagating |
|
|
| `mergeschool.core.alignment.apply_head_perms` permutes the query and output projections but not |
| `k_proj`/`v_proj`. That is exact for MHA and MQA but **not for grouped-query attention**: on |
| `Llama-3.1-8B` a flat head permutation changes the logits by **relative 1.115** — it destroys the |
| model. The group-respecting action implemented here (`code/gmap.py`) is exact to **9.4e-07**. Any |
| merge study that accepts a flat head permutation on a GQA model is silently corrupting its merges. |
|
|
