SCOPE

Research checkpoint repository for the SCOPE-KG project and its controlled baseline comparisons on ScienceQA and a documented Text–Image resplit of MultiModalQA.

Experiments are in progress. A checkpoint is not evidence that SCOPE outperforms MIRA or that the paper is complete. Validation and test results are distinguished in each run manifest. Checkpoint selection uses validation only; test scores must not guide subsequent training or model selection.

Uploads contain project-trained reader heads/adapters and, when applicable, project-trained retriever/selector weights, with configuration and SHA-256 provenance. They are not standalone Qwen models unless explicitly described as such. Base-model checkpoints, dataset images/text and third-party papers are not included in this initial release.

Each released run records the model, dataset, seed, selected epoch, evaluation scope and any required upstream dependencies. Controlled adaptations use a shared Qwen3-8B backbone and should not be interpreted as exact reproduction of every original paper's native setup.

The repository is public at the project owner's request. No credentials or private environment files belong in this repository.

On 10 October 2026 the owner requested a fresh experiment restart. Historical artifacts under earlier run paths are not results of the new run. The new run trains a fresh shared adapter and reruns every method; fixed data preprocessing may be reused after hash verification.