vantuan5644/coin-cl-qwen2.5-vl-7b-baselines

CoIN continual-instruction-tuning baselines — Qwen2.5-VL-7B

LoRA adapters for continual-learning baselines trained sequentially on the 8-task CoIN benchmark, in this order:

ScienceQA → TextVQA → ImageNet → GQA → VizWiz → Grounding → VQAv2 → OCR-VQA

Each <method>/Task<N>_<Name>/ holds the adapter after training task N, so the forward-transfer matrix evaluates checkpoint N on every task K<=N.

Check the tasks column below: a run with fewer than 8 is either a partial sequence (a hyperparameter probe or an ablation) or a run that trains once instead of once per task. The sections after the table cover the second kind.

Contents

run	size	tasks
clmoe	14G	8
disco	7.4G	8
disco_dko	7.4G	8
disco_s1	7.4G	8
disco_s2	7.4G	8
ewc	22G	8
ewc_cal	22G	8
hide	7.4G	8
joint_static	925M	1
modal_prompt	4.0G	8
moelora	7.4G	8
olora	7.4G	8
same	9.6G	8
sefe	25G	8
seqft	7.3G	8
smolora	7.4G	8

joint_static

The multi-task upper bound, not a continual run, so its tasks count is 1. One LoRA adapter is trained for one epoch over the union of the eight CoIN training sets (573,840 examples, built by scripts/build_joint_trainset.py), with the same rank, learning rate and effective batch as seqft. The only difference from seqft is that the data arrives mixed instead of in sequence. It is stored as Task1_Joint and scored against the eight canonical per-task test configs, giving one row of eight cells and no forgetting to measure.