| --- |
| license: apache-2.0 |
| tags: [model-merging, mode-connectivity, mergeability, pythia] |
| --- |
| |
| # crossarch-accuracy — downstream accuracy for the ~1B merge results |
|
|
| Companion to [`Mergeability-2/crossarch-1b-diagnostics`](https://huggingface.co/datasets/Mergeability-2/crossarch-1b-diagnostics), |
| which measured 106 Pythia-1.4B / Zh-Pythia-1.4B checkpoint-merge pairs, 1 native cross-model |
| merge and 20 transport merges **entirely in nats/token**. This repo adds the **accuracy** axis: |
| does any of it produce a *usable* model? |
|
|
| **Headline question (Figure 1).** The diagnostics found that checkpoint pairs within about half |
| a decade of training are linearly mode-connected (median LMC barrier −0.033 nats) while pairs |
| beyond 1.5 decades are not (+0.380). If Δaccuracy of the merge against the *better parent* |
| crosses zero at the same separation, the barrier — computable from step numbers plus one cheap |
| NLL sweep — is a free predictor of merge usability. |
|
|
| ## Benchmarks |
|
|
| Direct loglikelihood scorers, lm-eval-harness formats, validated against published |
| `EleutherAI/pythia-1.4b` numbers: |
|
|
| | task | ours | published | chance | |
| |---|---|---|---| |
| | SciQ | 0.869 | 0.865 | 0.25 | |
| | ARC-Easy | 0.605 | 0.616 | 0.25 | |
| | PIQA | 0.709 | 0.708 | 0.50 | |
| | LAMBADA (OpenAI) | 0.627 | 0.616 | 0.00 | |
| | Belebele (eng_Latn) | 0.220 | — | 0.25 (**at chance**) | |
| |
| `ACC` in the tables is the macro-average of the four tasks the 1.4B parents are above chance on |
| (ARC-Easy, SciQ, PIQA, LAMBADA); its pooled chance level is 0.25. Belebele is reported but is |
| at chance at this scale and must not be read as signal. |
| |
| ## Files |
| |
| - `results/table1_checkpoint_accuracy.csv` — Table 1, one row per checkpoint pair |
| - `results/acc_ledger.jsonl` — raw per-model per-task accuracies (resumable ledger) |
| - `results/summary.json` — crossover estimates and correlations with permutation nulls |
| - `figures/fig1_crossover_*.png` — Figure 1 |
| - `figures/fig2_cka_vs_gain.png` — Figure 2 |
| - `scripts/` — the harness, verbatim |
|
|
| Work in progress; published incrementally as cells land. |
|
|
