Simplify the standalone encoder card

README.md

@@ -1,5 +1,7 @@
 ---
 license: cc-by-4.0
+datasets:
+- risenyard/egms-qa-dataset
 tags:
 - remote-sensing
 - insar
@@ -8,30 +10,42 @@ tags:
 - egms
 ---
 
-# EGMS Encoder 4.3 for EGMS-QA
+# EGMS-QA Encoder
 
-Frozen EGMS Encoder 4.3 spatio-temporal checkpoint for EGMS-QA, a natural-language
-question-answering system over European Ground Motion Service persistent-
-scatterer displacement time series.
+A frozen spatio-temporal encoder for persistent-scatterer displacement time
+series from the European Ground Motion Service (EGMS).
 
-The encoder maps a variable-size 7 km tile to 65 tokens: one tile-summary token
-plus 64 spatial-cell tokens on an 8×8 grid, each with dimension 256. It was
-trained by synchronized-block masked reconstruction and remains frozen during
-translator training.
+For each 7 km tile, the encoder converts a variable number of point histories
+into context-aware 256-dimensional point representations. Deterministic 8×8
+spatial pooling then produces 65 EGMS tokens: one tile-summary token and 64
+cell tokens.
 
-Code: [`risenyard/egms-qa`](https://github.com/risenyard/egms-qa/tree/main/src/egms_encoder) ·
-Dataset: [`risenyard/egms-qa-dataset`](https://huggingface.co/datasets/risenyard/egms-qa-dataset) ·
-Translators: [`risenyard/egms-qa-translator`](https://huggingface.co/risenyard/egms-qa-translator)
+## Input and output
+
+| stage | shape |
+|---|---|
+| processed source tile | coordinates `[N,2]`, displacement series `[N,304]` |
+| encoder input | `[N,294]`, using the fixed slice `[8,302)` |
+| point representations | `[N,256]` |
+| pooled EGMS tokens | `[65,256]` plus a 65-element validity mask |
+
+The model was pretrained with synchronized-block masked reconstruction. Its
+weights remain frozen when training the language-model translator.
 
 ## Files
 
 | file | contents |
 |---|---|
-| `encoder.pt` | model state dictionary and embedded training arguments |
-| `normalization.json` | model-local mirror of the dataset's train-split mean, standard deviation, and residual statistics |
-| `args.json` | normalized public training and architecture configuration |
+| `encoder.pt` | encoder state dictionary and embedded architecture arguments |
+| `args.json` | readable architecture and training configuration |
+| `normalization.json` | input mean, standard deviation, and residual scale |
+
+Together these files contain the model-specific information needed to load the
+checkpoint. EGMS measurements are not included in this model repository.
 
-## Use with the structured dataset release
+## Use
+
+Install the code and download the checkpoint:
 
 ```bash
 git clone https://github.com/risenyard/egms-qa
@@ -40,41 +54,32 @@ pip install -e .
 
 hf download risenyard/egms-qa-encoder \
     --local-dir data/encoder/checkpoint
-hf download risenyard/egms-qa-dataset \
-    --repo-type dataset --local-dir release/egms-qa-dataset
-python -m egms_qa.release install \
-    --release-dir release/egms-qa-dataset --target-root .
-
-python -m egms_encoder.extract_tokens --output-dir outputs/tokens
-# -> outputs/tokens/encoder_tokens_10k.pt
 ```
 
-Most translator users can use the released `encoder_tokens_10k.pt` cache under
-`artifacts/representations/` in the dataset repository and skip encoder
-execution entirely.
+Extract tokens from an already prepared compatible NPZ tile collection:
 
-## Data and scope
+```bash
+python -m egms_encoder.extract_tokens \
+    --checkpoint data/encoder/checkpoint/encoder.pt \
+    --manifest path/to/split.parquet \
+    --data-config path/to/data_config.json \
+    --output-dir outputs/tokens
+```
 
-The model consumes the `[8,302)` slice of each processed 304-step source tile,
-giving 294 input steps. Coordinates are centred per tile and divided by the
-3,500 m half-width before coordinate embedding.
+The manifest and data config describe the user's tiles; they are not properties
+of the checkpoint. The released training data and reference manifest are
+available separately in
+[`risenyard/egms-qa-dataset`](https://huggingface.co/datasets/risenyard/egms-qa-dataset).
 
-The encoder supports the released 10,000-tile dataset and new tiles that
-already follow the EGMS-QA NPZ contract and released 304-step source axis. A
-user-prepared training corpus must also provide its own split manifest, data
-config, and train-fitted normalization. This release does not download official
-EGMS data, convert arbitrary EGMS ZIP/CSV products, or automatically select a
-time window and normalization for another reference period. Do not reuse the
-released `[8,302)` window blindly for a different EGMS product version.
+## Scope
 
-For release reproduction and standalone inference, the authoritative
-normalization file is `normalization.json` beside this checkpoint (SHA256
-`78f2ec35fbb8d2fdeb89d46fceb18257144af73b8d32bb506409a756f561b059`).
+The encoder expects data that already follow the EGMS-QA NPZ and time-axis
+contract. It does not download or convert official EGMS products. Its outputs
+describe observed deformation histories; they do not establish cause, forecast
+motion, or certify structural safety.
 
-The encoder represents observed deformation histories; it does not determine
-cause, forecast motion, certify structural safety, or replace expert InSAR and
-geotechnical assessment.
+Code: [`risenyard/egms-qa`](https://github.com/risenyard/egms-qa) ·
+Dataset: [`risenyard/egms-qa-dataset`](https://huggingface.co/datasets/risenyard/egms-qa-dataset) ·
+Translator: [`risenyard/egms-qa-translator`](https://huggingface.co/risenyard/egms-qa-translator)
 
-The checkpoint is CC-BY-4.0. Copernicus-derived source measurements are
-distributed separately in `risenyard/egms-qa-dataset` under the CLMS data
-policy and attribution conditions documented there.
+Released under CC-BY-4.0.