| # Cross-Backbone Adapter Swap — train on X, evaluate on Y |
|
|
| Evaluations: **240** rows (2 model sizes x 4 tasks x 3 eval backbones x (3 train backbones x 3 seeds + no-adapter)) |
|
|
| `penalty = matched accuracy − swapped accuracy`, computed **within a fixed evaluation backbone**, so it isolates adapter transfer from backbone quality. |
|
|
| ## Reference floor — backbone with NO adapter |
|
|
| | model_size | bf16 | int8 | nf4 | |
| |:-------------|-------:|-------:|-------:| |
| | 1B | 0.5196 | 0.5045 | 0.4962 | |
| | 3B | 0.5049 | 0.5050 | 0.5047 | |
| |
| ## Llama-3.2-1B |
| |
| Accuracy (rows = TRAIN backbone, cols = EVAL backbone): |
| |
| | train_bitwidth | bf16 | int8 | nf4 | |
| |:-----------------|-------:|-------:|-------:| |
| | bf16 | 0.8251 | 0.8140 | 0.7928 | |
| | int8 | 0.8216 | 0.8203 | 0.7894 | |
| | nf4 | 0.8118 | 0.8099 | 0.8272 | |
|
|
| Transfer penalty in percentage points (positive = swapping hurts): |
|
|
| | train_bitwidth | bf16 | int8 | nf4 | |
| |:-----------------|-------:|-------:|------:| |
| | bf16 | +0.00 | +0.63 | +3.44 | |
| | int8 | +0.35 | +0.00 | +3.77 | |
| | nf4 | +1.33 | +1.05 | +0.00 | |
| |
| - mean off-diagonal penalty: **+1.76 pp** (worst +3.77 pp, best +0.35 pp) |
| |
| ## Llama-3.2-3B |
| |
| Accuracy (rows = TRAIN backbone, cols = EVAL backbone): |
| |
| | train_bitwidth | bf16 | int8 | nf4 | |
| |:-----------------|-------:|-------:|-------:| |
| | bf16 | 0.8940 | 0.8931 | 0.8942 | |
| | int8 | 0.8977 | 0.8945 | 0.8935 | |
| | nf4 | 0.8928 | 0.8906 | 0.8960 | |
|
|
| Transfer penalty in percentage points (positive = swapping hurts): |
|
|
| | train_bitwidth | bf16 | int8 | nf4 | |
| |:-----------------|-------:|-------:|------:| |
| | bf16 | +0.00 | +0.14 | +0.18 | |
| | int8 | -0.37 | +0.00 | +0.24 | |
| | nf4 | +0.12 | +0.39 | +0.00 | |
| |
| - mean off-diagonal penalty: **+0.12 pp** (worst +0.39 pp, best -0.37 pp) |
| |
| ## Per-task penalty |
| |
| | | ('bf16', 'int8') | ('bf16', 'nf4') | ('int8', 'bf16') | ('int8', 'nf4') | ('nf4', 'bf16') | ('nf4', 'int8') | |
| |:---------------|-------------------:|------------------:|-------------------:|------------------:|------------------:|------------------:| |
| | ('1B', 'mnli') | +0.17 | +1.55 | -0.07 | +1.45 | +0.85 | +0.60 | |
| | ('1B', 'qqp') | +0.50 | +0.37 | -0.27 | +0.45 | -0.12 | +0.22 | |
| | ('1B', 'rte') | +1.93 | +11.79 | +1.68 | +13.12 | +4.57 | +3.25 | |
| | ('1B', 'sst2') | -0.08 | +0.04 | +0.04 | +0.08 | +0.00 | +0.11 | |
| | ('3B', 'mnli') | -0.03 | +0.03 | -0.17 | +0.00 | +0.53 | +0.43 | |
| | ('3B', 'qqp') | +0.67 | +0.62 | -0.32 | -0.07 | +0.33 | +0.75 | |
| | ('3B', 'rte') | -0.36 | -0.24 | -0.72 | +0.72 | -0.72 | -0.12 | |
| | ('3B', 'sst2') | +0.31 | +0.31 | -0.27 | +0.31 | +0.34 | +0.50 | |
| |
| ## Interpretation |
| |
| **Degenerate cells excluded from the headline** (mean accuracy within 10 pp of the majority-class floor — the adapter barely learned, so its 'transfer penalty' is mostly label noise): |
| |
| | model_size | task | acc | floor | margin | |
| |:-------------|:-------|-------:|--------:|---------:| |
| | 1B | rte | 0.5681 | 0.5271 | 0.0410 | |
|
|
| Mean off-diagonal transfer penalty (pp): |
|
|
| | model_size | all_tasks_pp | excluding_degenerate_pp | |
| |:-------------|---------------:|--------------------------:| |
| | 1B | +1.76 | +0.33 | |
| | 3B | +0.12 | +0.12 | |
| |
| - Excluding degenerate cells, the mean transfer penalty is **+0.21 pp**. |
| |
| - **Adapters transfer between backbones essentially for free.** Combined with the geometry result (cross-bit cosine 0.45–0.66, sign agreement ~0.65), this says the bit-width-induced rotation is **functionally redundant**: adapters trained against different backbones take measurably different routes to the same behaviour. Different solution, same function — evidence that the rank-16 solution space is highly degenerate. |
| |
| - Residual structure by **deployment** backbone (pp): 1B/eval-on-bf16 +0.07; 1B/eval-on-int8 +0.25; 1B/eval-on-nf4 +0.66; 3B/eval-on-bf16 -0.12; 3B/eval-on-int8 +0.27; 3B/eval-on-nf4 +0.21 |
| |
| - The small residual is concentrated on *deploying onto* NF4, not on *training on* it — consistent with NF4 being the most perturbed backbone, but the magnitude is under 1 pp. |
| |
| > ⚠️ Scope: this swaps ONE adapter at a time. It shows an individual adapter is robust to a backbone change; it does not by itself establish that MERGED adapters are, since merging compounds direction and sign disagreement across several adapters at once. |
| |
