londonaicentre · atriaybagur · Aug 28, 2026 · Aug 27, 2026 · Aug 27, 2026 · Aug 27, 2026
diff --git a/.gitignore b/.gitignore
@@ -456,16 +456,15 @@ flip-utils/dist/
 # version control (covered by .claude/* above, listed explicitly for clarity).
 .claude/worktrees/
 
-# fl-tutorials generated / uncommitted artifacts (data/, the spleen-seg uv.lock), scoped under the
-# per-backend tutorial trees (fl-tutorials/<backend>/) so a bare `data/` doesn't over-match
-# other data/ dirs in the monorepo:
-# - generated tutorial datasets (spleen/xray images + accession CSVs)
-# - tutorial uv.lock (tutorials intentionally don't pin a strict lockfile)
-fl-tutorials/nvflare/data/
-fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/uv.lock
+# fl-tutorials generated / uncommitted artifacts, scoped (not a bare `data/`, which would
+# over-match other data/ dirs in the monorepo):
+# - the shared tutorial-dataset root both backends' harnesses read (fl-tutorials/datasets/
+#   downloads land here: spleen/xray/arkplus images + accession CSVs + the HF cache)
+# - the dataset tooling's uv.lock (intentionally not pinned to a strict lockfile)
+fl-tutorials/data/
+fl-tutorials/datasets/spleen/uv.lock
 fl-services/flower/provision/creds/
 fl-services/flower/runs/
-fl-tutorials/flower/data/
 
 # Round-timing extraction output (scripts/fl_round_metrics/ extractors and the
 # tutorials' `make round-metrics` write

diff --git a/AGENTS.md b/AGENTS.md
@@ -16,7 +16,7 @@ FLIP/
 ├── flip-utils/         # FLIP Python library (pip-installable flip-utils)
 ├── fl-services/        # FL Docker services + network provisioning, per backend (Makefile owns build/provision/up/down/submit; flower also up-secure): fl-services/nvflare/{fl-base,fl-server,fl-client,fl-api-base, provision/{net-*_project_*.yml, scripts/, workspace-{dev,stag,prod}/ gitignored}}, fl-services/flower/{fl-base,superlink,supernode,fl-api-flower, provision/{scripts/, creds/ gitignored}} (#622)
 ├── fl-apps/            # FL app templates per backend: fl-apps/nvflare/{standard,evaluation,diffusion_model,fed_opt} (all Client-API), fl-apps/flower/{standard,evaluation} + check_required_files.sh (cross-backend CI validator at root)
-├── fl-tutorials/       # FL tutorials per backend (all NVFLARE ones are Client-API apps): fl-tutorials/nvflare/{image_*}, fl-tutorials/flower/{xray_classification,3d_spleen_segmentation*} (root Makefile forwards by FL_BACKEND); xray classification, spleen seg/eval, diffusion. Plus fl-tutorials/tests/ — CPU-only pytest over the tutorial transform chains (#871) plus a static `min_clients` wiring guard covering fl-apps/flower too, run by `make -C fl-tutorials test`
+├── fl-tutorials/       # FL tutorials per backend (all NVFLARE ones are Client-API apps): fl-tutorials/nvflare/{image_*}, fl-tutorials/flower/{xray_classification,3d_spleen_segmentation*} (root Makefile forwards by FL_BACKEND); xray classification, spleen seg/eval, diffusion. Shared dataset tooling in fl-tutorials/datasets/ (download/derive/enrich, single copy for both backends — the download-*-data + upload-spleen-labels targets), outputs in the shared gitignored fl-tutorials/data/. Plus fl-tutorials/tests/ — CPU-only pytest over the tutorial transform chains (#871) plus a static `min_clients` wiring guard covering fl-apps/flower too, run by `make -C fl-tutorials test`
 ├── map-apps/           # MONAI Application Package (MAP) templates for packaging FLIP-trained models for clinical deployment
 ├── trust/
 │   ├── trust-api/      # Trust API gateway (Python/FastAPI)
@@ -159,7 +159,7 @@ cause. (Older app copies die with torch's opaque `num_samples=0` instead.)
 
 `e2e_smoke` has a hook for exactly this: `--data-enrichment-cwd` + `--data-enrichment-cmd` run a shell command
 **between the image pull and training**, with `FLIP_PROJECT_ID` exported. Since FLIP#776 the spleen uploader is
-in-tree at `fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/utils/upload_spleen_labels_to_xnat.py`
+in-tree at `fl-tutorials/datasets/spleen/upload_spleen_labels_to_xnat.py` (the shared dataset-tooling tree)
 — **no private repo required**. It resolves each trust's XNAT project by `secondary_ID == <FLIP project_id>`,
 fetches the accession→MSD-case mapping at run time from the public `aicentreflip/trust-data` dataset
 (`omop-csv/<version>/spleen_project/image_occurrence.csv`, which also carries `source_trust`), and writes each
@@ -198,8 +198,8 @@ instead:
 cd flip-api && uv run python -m tests.e2e_smoke \
   --model-files-dir ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/app_files \
   --query-file ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/query.sql \
-  --data-enrichment-cwd ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation \
-  --data-enrichment-cmd 'uv run --no-project --with ../../../../flip-utils python utils/upload_spleen_labels_to_xnat.py --flip-project-id "$FLIP_PROJECT_ID" --labels-dir ../../data/spleen/images --xnat-url http://127.0.0.1:8104 --xnat-url http://127.0.0.1:8106'
+  --data-enrichment-cwd ../fl-tutorials/datasets/spleen \
+  --data-enrichment-cmd 'uv run --no-project --with ../../../flip-utils python upload_spleen_labels_to_xnat.py --flip-project-id "$FLIP_PROJECT_ID" --labels-dir ../../data/spleen/images --xnat-url http://127.0.0.1:8104 --xnat-url http://127.0.0.1:8106'
 ```
 
 (The `$`-escaping trap still applies to any hand-written `EXTRA_ARGS`: `make -C flip-api …` expands once, so
@@ -209,7 +209,7 @@ there. The in-Makefile targets handle this for you. Note that `e2e_smoke_spleen`
 tutorial (`../fl-tutorials/flower/3d_spleen_segmentation/{app,query.sql}`), not the NVFLARE paths shown in the
 direct-invocation example above — pair it with `FL_BACKEND=flower`, or invoke `tests.e2e_smoke` directly with
 the NVFLARE paths for the NVFLARE tutorial. The enrichment step itself is backend-agnostic and runs out of the
-NVFLARE tree either way.)
+shared `fl-tutorials/datasets/spleen/` tree either way.)
 
 Enrichment must land **after** the pull and after DICOM→NIfTI conversion; the hook's position guarantees that.
 The uploader derives each target filename from the converted `input_*.nii.gz`, so with no `NIFTI` resource it

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -16,7 +16,7 @@ FLIP/
 ├── flip-utils/         # FLIP Python library (pip-installable flip-utils)
 ├── fl-services/        # FL Docker services + network provisioning, per backend (Makefile owns build/provision/up/down/submit; flower also up-secure): fl-services/nvflare/{fl-base,fl-server,fl-client,fl-api-base, provision/{net-*_project_*.yml, scripts/, workspace-{dev,stag,prod}/ gitignored}}, fl-services/flower/{fl-base,superlink,supernode,fl-api-flower, provision/{scripts/, creds/ gitignored}} (#622)
 ├── fl-apps/            # FL app templates per backend: fl-apps/nvflare/{standard,evaluation,diffusion_model,fed_opt} (all Client-API), fl-apps/flower/{standard,evaluation} + check_required_files.sh (cross-backend CI validator at root)
-├── fl-tutorials/       # FL tutorials per backend (all NVFLARE ones are Client-API apps): fl-tutorials/nvflare/{image_*}, fl-tutorials/flower/{xray_classification,3d_spleen_segmentation*} (root Makefile forwards by FL_BACKEND); xray classification, spleen seg/eval, diffusion. Plus fl-tutorials/tests/ — CPU-only pytest over the tutorial transform chains (#871) plus a static `min_clients` wiring guard covering fl-apps/flower too, run by `make -C fl-tutorials test`
+├── fl-tutorials/       # FL tutorials per backend (all NVFLARE ones are Client-API apps): fl-tutorials/nvflare/{image_*}, fl-tutorials/flower/{xray_classification,3d_spleen_segmentation*} (root Makefile forwards by FL_BACKEND); xray classification, spleen seg/eval, diffusion. Shared dataset tooling in fl-tutorials/datasets/ (download/derive/enrich, single copy for both backends — the download-*-data + upload-spleen-labels targets), outputs in the shared gitignored fl-tutorials/data/. Plus fl-tutorials/tests/ — CPU-only pytest over the tutorial transform chains (#871) plus a static `min_clients` wiring guard covering fl-apps/flower too, run by `make -C fl-tutorials test`
 ├── map-apps/           # MONAI Application Package (MAP) templates for packaging FLIP-trained models for clinical deployment
 ├── trust/
 │   ├── trust-api/      # Trust API gateway (Python/FastAPI)
@@ -159,7 +159,7 @@ cause. (Older app copies die with torch's opaque `num_samples=0` instead.)
 
 `e2e_smoke` has a hook for exactly this: `--data-enrichment-cwd` + `--data-enrichment-cmd` run a shell command
 **between the image pull and training**, with `FLIP_PROJECT_ID` exported. Since FLIP#776 the spleen uploader is
-in-tree at `fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/utils/upload_spleen_labels_to_xnat.py`
+in-tree at `fl-tutorials/datasets/spleen/upload_spleen_labels_to_xnat.py` (the shared dataset-tooling tree)
 — **no private repo required**. It resolves each trust's XNAT project by `secondary_ID == <FLIP project_id>`,
 fetches the accession→MSD-case mapping at run time from the public `aicentreflip/trust-data` dataset
 (`omop-csv/<version>/spleen_project/image_occurrence.csv`, which also carries `source_trust`), and writes each
@@ -198,8 +198,8 @@ instead:
 cd flip-api && uv run python -m tests.e2e_smoke \
   --model-files-dir ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/app_files \
   --query-file ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation/query.sql \
-  --data-enrichment-cwd ../fl-tutorials/nvflare/image_segmentation/3d_spleen_segmentation \
-  --data-enrichment-cmd 'uv run --no-project --with ../../../../flip-utils python utils/upload_spleen_labels_to_xnat.py --flip-project-id "$FLIP_PROJECT_ID" --labels-dir ../../data/spleen/images --xnat-url http://127.0.0.1:8104 --xnat-url http://127.0.0.1:8106'
+  --data-enrichment-cwd ../fl-tutorials/datasets/spleen \
+  --data-enrichment-cmd 'uv run --no-project --with ../../../flip-utils python upload_spleen_labels_to_xnat.py --flip-project-id "$FLIP_PROJECT_ID" --labels-dir ../../data/spleen/images --xnat-url http://127.0.0.1:8104 --xnat-url http://127.0.0.1:8106'
 ```
 
 (The `$`-escaping trap still applies to any hand-written `EXTRA_ARGS`: `make -C flip-api …` expands once, so
@@ -209,7 +209,7 @@ there. The in-Makefile targets handle this for you. Note that `e2e_smoke_spleen`
 tutorial (`../fl-tutorials/flower/3d_spleen_segmentation/{app,query.sql}`), not the NVFLARE paths shown in the
 direct-invocation example above — pair it with `FL_BACKEND=flower`, or invoke `tests.e2e_smoke` directly with
 the NVFLARE paths for the NVFLARE tutorial. The enrichment step itself is backend-agnostic and runs out of the
-NVFLARE tree either way.)
+shared `fl-tutorials/datasets/spleen/` tree either way.)
 
 Enrichment must land **after** the pull and after DICOM→NIfTI conversion; the hook's position guarantees that.
 The uploader derives each target filename from the converted `input_*.nii.gz`, so with no `NIFTI` resource it

diff --git a/docs/source/user-guides/arkplus-fine-tuning.rst b/docs/source/user-guides/arkplus-fine-tuning.rst
@@ -104,7 +104,7 @@ published reference dataset on Hugging Face:
 
 This pulls the site-1 / site-2 training splits from
 `aicentreflip/tutorials-arkplus-cxr-classification <https://huggingface.co/datasets/aicentreflip/tutorials-arkplus-cxr-classification>`_
-and lays them out under ``fl-tutorials/nvflare/data/arkplus/`` (gitignored).
+and lays them out under ``fl-tutorials/data/arkplus/`` (gitignored).
 
 To use your own data instead, point the per-site ``.env.app`` values
 (``SITE{1,2}_IMAGES_DIR`` / ``SITE{1,2}_DATAFRAME``) at your directories. The trainer selects a

diff --git a/fl-services/flower/README.md b/fl-services/flower/README.md
@@ -155,10 +155,10 @@ trains on an empty dataset with no signal that the data is missing. So download
 point both variables at it when bringing the net up:
 
 ```bash
-make -C fl-tutorials download-xray-data FL_BACKEND=flower   # → fl-tutorials/flower/data/xrays_mini_300/
+make -C fl-tutorials download-xray-data FL_BACKEND=flower   # → fl-tutorials/data/xrays_mini_300/
 make -C fl-services/flower up \
-  DEV_IMAGES_DIR=$(pwd)/fl-tutorials/flower/data/xrays_mini_300/accession-resources \
-  DEV_DATAFRAME=$(pwd)/fl-tutorials/flower/data/xrays_mini_300/sample_get_dataframe_response.csv
+  DEV_IMAGES_DIR=$(pwd)/fl-tutorials/data/xrays_mini_300/accession-resources \
+  DEV_DATAFRAME=$(pwd)/fl-tutorials/data/xrays_mini_300/dataframe.csv
 ```
 
 (Run from the repo root; `data/` is gitignored. `up-secure` takes the same two variables.)

diff --git a/fl-tutorials/Makefile b/fl-tutorials/Makefile
@@ -9,22 +9,29 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Backend-aware entrypoint for the FL tutorial runners. The actual runner + harness live
-# per-backend under fl-tutorials/<backend>/ (fl-tutorials/nvflare/ and fl-tutorials/flower/). This thin
-# Makefile forwards every target to fl-tutorials/$(FL_BACKEND)/ so the documented entrypoint
+# Backend-aware entrypoint for the FL tutorial runners. The runner + harness live per-backend
+# under fl-tutorials/<backend>/ (fl-tutorials/nvflare/ and fl-tutorials/flower/); the shared
+# dataset tooling lives in fl-tutorials/datasets/ (outputs in the gitignored fl-tutorials/data/,
+# one download serving both backends). This thin Makefile forwards each target to the right
+# sub-make so the documented entrypoint
 #   make -C fl-tutorials <target> [TUTORIAL=... NUM_CASES=...]
 # keeps working and selects the backend via FL_BACKEND (default nvflare). Command-line
 # variables are forwarded to the sub-make automatically.
 FL_BACKEND ?= nvflare
 
-TUTORIAL_TARGETS := list-tutorials run-tutorial run-all-tutorials \
-	download-xray-data download-spleen-data upload-spleen-labels \
+TUTORIAL_TARGETS := list-tutorials run-tutorial run-all-tutorials
+# Dataset targets are backend-agnostic (shared data root, single copy in datasets/);
+# FL_BACKEND still selects a per-harness variant where one exists (the two spleen downloads).
+DATASET_TARGETS := download-xray-data download-spleen-data upload-spleen-labels \
 	download-arkplus-finetuning-data download-arkplus-eval-data
 
-.PHONY: $(TUTORIAL_TARGETS)
+.PHONY: $(TUTORIAL_TARGETS) $(DATASET_TARGETS)
 $(TUTORIAL_TARGETS):
 	@$(MAKE) --no-print-directory -C $(FL_BACKEND) $@
 
+$(DATASET_TARGETS):
+	@$(MAKE) --no-print-directory -C datasets $@ FL_BACKEND=$(FL_BACKEND)
+
 # ---------------------------------------------------------------------------
 # Lint / autoformat the tutorial sources. Runs ruff over the whole tree; ruff
 # auto-excludes the per-tutorial .venv/ trees and resolves config per-file

diff --git a/fl-tutorials/README.md b/fl-tutorials/README.md
@@ -47,9 +47,10 @@ make -C fl-tutorials run-tutorial TUTORIAL=xray_classification
 built `flare-fl-base` image and fail without it. The spleen examples take `download-spleen-data` in place of
 `download-xray-data`.
 
-The simulator requires Docker, and GPU-backed examples require the NVIDIA Container Toolkit. Dataset downloads and
-generated runs are kept in gitignored backend data/output directories. Run `make -C fl-tutorials run-all-tutorials`
-only when you intentionally want the full, heavyweight suite.
+The simulator requires Docker, and GPU-backed examples require the NVIDIA Container Toolkit. Dataset tooling is
+shared across backends in [`datasets/`](datasets/) and downloads land in the shared gitignored `data/` root, so
+one download serves both backends; generated runs stay in gitignored per-backend output directories. Run
+`make -C fl-tutorials run-all-tutorials` only when you intentionally want the full, heavyweight suite.
 
 ## Check the sources without a GPU