EGR-MED

# EGR-MED Dataset and Cache Release EGR-MED is the cache-native data release for evidence-guided clinical image forgery detection, localization, rationale SFT, and EG-GRPO. > The repository intentionally does not expose one image per file. RGB images, expert-overlay panels, and DIRE maps are stored in aligned binary shards with JSONL byte-range indices. ModelScope's generic dataset preview therefore does not apply; use the released EGR-Med loader and validation script. ## Code Training and evaluation code: <https://github.com/Max970304/egr-med> ## Download ```bash pip install modelscope modelscope download --dataset Max0304/EGR-MED --local_dir /path/to/EGR-MED ``` The complete release is approximately 1.49 TB (decimal). Use ModelScope's `--include` option when only a subset is needed. ## Layout ```text EGR-MED/ ├── release_manifest.json ├── forgery_dataset_canonical512/ │ ├── canonical512_manifest.json │ ├── canonical512_sizes.jsonl │ └── splits/egrmed_seed2026/ ├── cache/ │ ├── rgb/rgb512_square_uint8_canonical512/ │ ├── overlay_panels/qwen3vl_full_heatmap_abs256_alpha045_canonical512/ │ └── dire_sharded_canonical512/openai_guided_diffusion_512_u8_sharded/ └── runs/ ├── step2_expert_evidence_full32_no_tokens_cacheorder_singlefile_emed_q90q99_20260704/ └── step4_qwen37plus_nothink_full_heatmap_realtime_20260706_qwen36fallback_fused/sft_targets/ ``` ## Workflow requirements | Workflow | Required prefixes | |---|---| | Overlay SFT | `forgery_dataset_canonical512`, RGB cache, overlay cache, SFT targets | | EG-GRPO | Overlay-SFT assets plus the latest q90--q99 Step-2 evidence | | Stage-1 expert retraining | canonical metadata/splits, RGB cache, and DIRE cache for `E_gen` | The DIRE cache is not needed when only reproducing Overlay SFT and EG-GRPO. Teacher-generated rationale targets are released for `train` and `val` only. Test evaluation is assembled from the canonical metadata and aligned RGB, overlay, and Step-2 assets and does not use a teacher rationale. ## Validation After cloning the code repository: ```bash python scripts/validate_release_data.py --data-root /path/to/EGR-MED --scope all ``` The SFT and GRPO exporters then perform row-level alignment checks over sample keys, labels, bounding boxes, grid size, cache offsets, and split lengths. ## Task contract - Authentic images: verdict `authentic`, bounding box `none`. - F1--F4 localized manipulations: verdict `forged`, one normalized manipulation box. - F5 full-image synthesis: verdict `forged`, bounding box `none`. - The mainline expert evidence grid is fixed at `32 x 32`. ## Data responsibility This release is intended for research on clinical-image integrity. Users are responsible for complying with the licenses and usage conditions of all upstream clinical datasets and foundation models, and for avoiding attempts to identify individuals or deploy the system as a substitute for clinical judgment.