Add files using upload-large-folder tool

README.md

@@ -0,0 +1,52 @@
+---
+license: cc-by-4.0
+tags:
+- remote-sensing
+- insar
+- ground-motion
+- time-series
+- egms
+---
+
+# EGMS-QA · Encoder
+
+Frozen spatio-temporal encoder and its 10,000-tile token cache for **EGMS-QA**,
+a natural-language question-answering system over
+[European Ground Motion Service](https://egms.land.copernicus.eu/) (EGMS)
+persistent-scatterer displacement time series.
+
+The encoder maps a variable-size 7 km tile of point displacement histories to a
+fixed **65-token** representation (1 tile summary + 8×8 spatial cells, 256-d),
+trained self-supervised with masked reconstruction. It is frozen downstream.
+
+Code: https://github.com/risenyard/egms-qa ·
+Companion repos: [`risenyard/egms-qa-qa_construction`](https://huggingface.co/datasets/risenyard/egms-qa-qa_construction) ·
+[`risenyard/egms-qa-translator`](https://huggingface.co/risenyard/egms-qa-translator)
+
+## Files
+
+| file | what it is |
+|---|---|
+| `encoder.pt` | frozen encoder checkpoint (`model` state dict + training args) |
+| `encoder_tokens_10k.pt` | token cache for the 10k tiles: `spatial_tokens [10000,65,256]` (X), `token_mask [10000,65]` (m, empty cells = 0), plus `tile_ids`, `splits`, `point_count_per_bin`, `n_points_per_tile`, `metadata` |
+| `encoder_tokens_10k_metadata.json` | extraction metadata |
+| `manifest/split.parquet` | 10k-tile train/val/test split manifest |
+| `manifest/data_config.json` | data configuration reference |
+
+## Use
+
+```python
+from huggingface_hub import hf_hub_download
+import torch
+
+tok = torch.load(hf_hub_download("risenyard/egms-qa-encoder", "encoder_tokens_10k.pt"),
+                 map_location="cpu", weights_only=False)
+X, m = tok["spatial_tokens"], tok["token_mask"]   # [10000,65,256], [10000,65]
+```
+
+Re-extracting tokens from raw tiles, or loading the checkpoint architecture,
+needs the model code in the GitHub repo (`egms_encoder`). Place these files at
+`data/encoder/…` of a checkout to reproduce end to end (see the repo README).
+
+The raw EGMS Level-3 product is Copernicus data and is not redistributed here.
+Licensed CC-BY-4.0.

encoder.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdb0bd2437c063a79c871eb54445dbee6fb82bd9e20186768fa5b947ffd9d474
+size 79442395

encoder_tokens_10k.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c95bb465bf4a948d8858bcdf51a0ce851e6ba9e849a8db2772fca4f49b14312a
+size 669285010

encoder_tokens_10k_metadata.json

@@ -0,0 +1,77 @@
+{
+  "encoder_checkpoint": "data/encoder/checkpoint/encoder.pt",
+  "coord_scale": 3500.0,
+  "encoder_args": {
+    "v4_config": "data/processed/v4/v4_data_config.json",
+    "v4_normalization": "data/processed/v4/v4_normalization.json",
+    "output_dir": "outputs/encoder_pretrain",
+    "tile_size": 7000.0,
+    "min_tile_points": 200,
+    "max_tile_points": 4096,
+    "tiles_per_batch": 8,
+    "d_model": 256,
+    "num_layers": 6,
+    "num_heads": 8,
+    "dropout": 0.1,
+    "input_length": 294,
+    "mask_ratio": 0.3,
+    "mask_strategy": "block",
+    "sync_mask": true,
+    "mask_schedule": "fixed",
+    "eval_mask_ratio": 0.3,
+    "model_version": "v3_3",
+    "patch_size": 8,
+    "temporal_layers": 2,
+    "temporal_heads": 4,
+    "residual_loss_weight": 1.0,
+    "residual_consistency_weight": 0.1,
+    "residual_head_mode": "additive",
+    "coord_scale": 3500.0,
+    "residual_head_lr": 0.0001,
+    "init_from_checkpoint": null,
+    "point_sampling": "residual_weighted",
+    "residual_sampling_alpha": 0.75,
+    "max_steps": 150000,
+    "duration_hours": null,
+    "lr": 0.0004,
+    "min_lr": 3e-05,
+    "lr_scheduler": "cosine",
+    "scheduler_total_steps": 150000,
+    "warmup_steps": 1000,
+    "weight_decay": 0.01,
+    "precision": "bf16",
+    "device": "cuda:0",
+    "seed": 42,
+    "split_strategy": "random",
+    "val_fraction": 0.1,
+    "test_fraction": 0.1,
+    "stratify_bins": 3,
+    "val_batches": 16,
+    "val_every_steps": 1000,
+    "val_seed": 1729,
+    "checkpoint_every_steps": 5000,
+    "log_every_steps": 200,
+    "train_window_steps": 1000,
+    "resume_from": null
+  },
+  "manifest_path": "data/encoder/manifest/split.parquet",
+  "data_config_path": "data/encoder/manifest/data_config.json",
+  "normalizer_mean": -2.6937686788514412,
+  "normalizer_std": 5.697023090760733,
+  "tile_size": 7000.0,
+  "grid_size": 8,
+  "n_tokens": 65,
+  "n_tiles": 10000,
+  "d_model": 256,
+  "input_length": 294,
+  "time_window_start": 8,
+  "time_window_end": 302,
+  "token_layout": "index 0 = CLS, 1..64 = 8x8 bins row-major",
+  "extraction_script": "src/egms_encoder/extract_tokens.py",
+  "extraction_date_utc": "2026-07-01T00:26:39Z",
+  "bin_occupancy_mean": 0.8588296875,
+  "bin_occupancy_median": 0.9375,
+  "n_points_per_tile_mean": 2018.0485,
+  "n_points_per_tile_min": 200,
+  "n_points_per_tile_max": 4900
+}

manifest/data_config.json

@@ -0,0 +1,53 @@
+{
+  "version": "v4",
+  "created": "2026-05-27",
+  "source": {
+    "tiles_dir": "data/processed/tiles_l3_u_2019_2023_europe",
+    "total_tiles_available": 131281
+  },
+  "filter": {
+    "max_nan_mean": 0.02,
+    "require_first_valid": 0,
+    "require_last_valid_min": 301,
+    "min_points": 200,
+    "pool_size": 83323,
+    "pool_total_points": 101340689
+  },
+  "time_window": {
+    "t_start": 8,
+    "t_end": 302,
+    "input_length": 294,
+    "rationale": "Pool tiles are 100% NaN-free within [t=8, t=302). Earlier steps (t=1,4-7) and tail steps (t=302,303) are systematically missing in 10-21% of pool tiles and are trimmed. Verified by scripts/v4_verify_trim_zero_nan.py on full 83K pool: 0 residual NaN.",
+    "compared_to_v33": "V3.3 used 302 steps (t=0..301 of the no_tail parquet). V4 uses 294 steps (t=8..301 of the raw npz). V4 starts ~6 weeks later in 2019."
+  },
+  "geographic_sampling": {
+    "method": "grid_max_npoints_with_round_robin",
+    "grid_m": 30000,
+    "target_tiles": 10000,
+    "sampled_tiles": 10000,
+    "occupied_cells": 4142,
+    "cells_in_sample": 4142,
+    "rounds_used": 3,
+    "total_points_sampled": 20180485,
+    "seed": 42
+  },
+  "files": {
+    "pool_candidates": "data/processed/v4/v4_pool_candidates.parquet",
+    "sample_10k": "data/processed/v4/v4_sample_10k.parquet",
+    "split": "data/processed/v4/v4_split.json",
+    "nan_profiles_source": "outputs/archive/V4/nan_pattern_analysis_20260527_145457/v4_nan_profiles.parquet"
+  },
+  "split": {
+    "method": "plain_random_tile_level",
+    "ratios": {"train": 0.80, "val": 0.10, "test": 0.10},
+    "counts": {"train": 8000, "val": 1000, "test": 1000},
+    "seed": 1729,
+    "n_grid_cells_per_split": {"train": 4022, "val": 925, "test": 923, "total_cells": 4142},
+    "backup_spatial_blocked": "data/processed/v4/v4_split_spatial_blocked.{json,parquet}.bak"
+  },
+  "tile_field_layout": {
+    "feature_columns_count": 10,
+    "expected_layout": "[easting, northing, height, rmse, mean_velocity, mean_velocity_std, acceleration, acceleration_std, seasonality, seasonality_std, time_series(294)]",
+    "note": "LazyTileStore must construct rows in this order from per-tile npz to remain compatible with V3.3 trainer expectations."
+  }
+}

manifest/split.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfb92ef7604d6ab18449f24a10d519985030c87cf6bf380dbc338ed5c3d91784
+size 434406