CG-CWRA — Cross-window Graph Attention Adapter (WTQ generative QA)
Frozen-TextMonkey adapter checkpoints for generative table QA on
WikiTableQuestions (WTQ). The adapter is inserted between the ViT window
features and the visual–language connector (downresampler); TextMonkey's
backbone and connector are frozen (connector is trained only in the
conn_only control variants).
Originally produced for an ablation study of cross-window modelling;
conn_only currently outperforms the CG variants, so these weights are
mainly useful for reproducing the study rather than as a drop-in SOTA model.
Variants
| File | ViT window interaction | CG adapter | Trainable | best val EM |
|---|---|---|---|---|
cgcwra_best.pt |
ON | 285M | 299.5M | 0.2455 (ep5) |
cgcwra_pure_best.pt |
OFF | 285M | 299.5M | 0.2436 (ep9) |
conn_only_best.pt |
ON | — | 13.6M | 0.2851 (ep7) |
conn_only_nointer_best.pt |
OFF | — | 13.6M | 0.2673 (ep7) |
Test metrics (WTQ, table-disjoint split_v2, seed 7, greedy max_new_tokens=32)
| Variant | EM | ANLS |
|---|---|---|
| TextMonkey zero-shot | 0.2279 | 0.1664 |
conn_only |
0.2688 | 0.1894 |
cgcwra |
0.2408 | 0.1742 |
cgcwra_pure |
0.2128 | 0.1529 |
conn_only_nointer |
pending | pending |
- EM = normalized exact match; ANLS = normalized Levenshtein similarity (τ=0.5).
- Single seed: differences below ≈0.02 EM are within run-to-run noise.
- No claim is made that CG supersedes or replaces TextMonkey's ViT window interaction.
Files
cgcwra_best.pt / cgcwra_pure_best.pt / conn_only_best.pt / conn_only_nointer_best.pt
histories/<variant>_history.json # 10-epoch train/val curves
qa_experiment_report_test.json # protocol + per-variant metrics + equivalence checks
Each *.pt contains adapter (null for conn_only*) and downresampler
state dicts plus epoch / val_em / seed metadata.
Protocol
- Input: 896×896 image → 4 windows (448×448) + 448 global; 768 visual tokens; prompt
<img></img> {question} Answer:. - Training: 10 epochs, batch 6, lr 1e-4 cosine (5% warmup), seed 7; checkpoint = best val-EM epoch.
- Data: WTQ, table-disjoint split (train 4401 / val 505 / test 1931).
Caveats
- Best-epoch weights on an overfitting curve (val loss rises after ep5–7); fine-tune further with regularization if reusing.
- Base TextMonkey / Qwen-VL weights are not redistributed here.