Fix released weights: publish trained final state (issue #2)

The previous shards contained an EMA state that was never updated during training (initialization weights) due to a key-prefix mismatch in update_ema under torch.compile + DeepSpeed. This upload replaces them with the final 300K-step training weights, verified to load strictly into Flux2 and to denoise (delta_rms > 0). Adds SHA256SUMS.

README.md

@@ -36,12 +36,19 @@ exclusively on EO data. GeoCore-9B is a Flow Matching Diffusion Transformer that
 generation on text **and** continuous geospatial metadata — ground sample distance (GSD), latitude
 and longitude.
 
+> [!NOTE]
+> **2026-08-04 — corrected weights.** The initial upload accidentally contained an EMA state that
+> was never updated during training (pure initialization weights — see
+> [GitHub issue #2](https://github.com/KAIST-VICLab/GeoCore-9B/issues/2)). The current files are the
+> final 300K-step training weights, verified to load strictly and denoise correctly. If you
+> downloaded the weights before this date, please re-download and check them against `SHA256SUMS`.
+
 ## Model details
 
 | | |
 |---|---|
 | Parameters | 9.24 B |
-| Weights | EMA, bfloat16, sharded safetensors |
+| Weights | final training weights (non-EMA), bfloat16, sharded safetensors |
 | Architecture | Flow Matching DiT — 8 double-stream + 24 single-stream blocks, hidden 4096, 32 heads |
 | Conditioning | CLIP + T5 text embeddings, GSD, latitude, longitude |
 | Training data | [Git-10M](https://huggingface.co/datasets/lcybuaa/Git-10M) |

SHA256SUMS

@@ -0,0 +1,6 @@
+37867b56a02748c67fe64f60118b6b6cd6966653026ba101abd1bf849e4f4b37  model-00001-of-00004.safetensors
+561831477a146da51524beacbd28e4e4db9c3d66318c98388823e641e920a956  model-00002-of-00004.safetensors
+60107ab28114906426b4187bbca9fda59af121318cb39543a90697b7a7ecef9a  model-00003-of-00004.safetensors
+7cef9bd84c652cb25d0868d0f06d3cc4279d8a066db219f9c5f3fb4c7ac7288a  model-00004-of-00004.safetensors
+6af0afb014d208f28e04804cb0bbe8bdf3b52a8490cf3610e001b85c17fa8b65  model.safetensors.index.json
+5b289da77aa6062d83ddb87c81b450e452e75f488401329cabbc0890ee2805b3  config.json

config.json

@@ -18,6 +18,6 @@
   "architecture": "Flux2",
   "model_size": "9b",
   "torch_dtype": "bf16",
-  "weights": "ema",
+  "weights": "model",
   "training_steps": 300000
 }

model-00001-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a7057a4d09362556f8ff2f0bf2a15561adb5f24449f124105896055881f9c2a6
+oid sha256:37867b56a02748c67fe64f60118b6b6cd6966653026ba101abd1bf849e4f4b37
 size 4948246312

model-00002-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5d6beb4d01c4773f8b2b3cd333ad3a92dc06eb933ad29963aaaf32e1c5d0868d
+oid sha256:561831477a146da51524beacbd28e4e4db9c3d66318c98388823e641e920a956
 size 4966067040

model-00003-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d90c65dd8ff6de3baddb7446f5ba6fa9ab721e8f2b12a42d1e462b9eb8530a3e
+oid sha256:60107ab28114906426b4187bbca9fda59af121318cb39543a90697b7a7ecef9a
 size 4932512840

model-00004-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8127a1d0bbc535562920e12debca65a0eb91434fd91026f33ef755ec8e469d56
+oid sha256:7cef9bd84c652cb25d0868d0f06d3cc4279d8a066db219f9c5f3fb4c7ac7288a
 size 3624934504