Release 0.1.0

README.md

@@ -1,3 +1,23 @@
 ---
-license: mit
+tags:
+  - porefm
+  - pytorch
+  - nanopore
+  - dna
+  - pretrained
 ---
+
+# PoreFM-DNA
+
+Pretrained PoreFM model for Oxford Nanopore DNA signals.
+
+Model version: **0.1.0**. Chemistry: **R10.4.1 (E8.2, 400 bps)**. Embedding dimension: **384**.
+
+| File | Contents |
+|---|---|
+| `backbone.pt` | Backbone for embeddings and downstream tasks. |
+| `pretraining.pt` | Backbone, pretraining objective heads, and EMA target encoder; optimizer and scheduler states are excluded. |
+
+Model configuration and preprocessing are embedded in the checkpoints. Setup and usage are documented in the PoreFM code repository.
+
+`manifest.json` records artifact identities and tensor fingerprints. `SHA256SUMS` verifies checkpoint downloads. Tensor fingerprints are independent of descriptive metadata; file hashes identify exact bytes. Documentation-only changes leave the model version unchanged.

SHA256SUMS

@@ -0,0 +1,2 @@
+4367c9a7ef291108a4bfcf6ae8452723c06bb68aa89f715ad4b8115b560986ac  backbone.pt
+d120e0e9abba56fe8ec62cf8549e5fb969c84321982a0d55d4113fe670fdcf2f  pretraining.pt

backbone.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4367c9a7ef291108a4bfcf6ae8452723c06bb68aa89f715ad4b8115b560986ac
+size 90747343

manifest.json

@@ -0,0 +1,43 @@
+{
+  "schema_version": 1,
+  "model_name": "PoreFM-DNA",
+  "model_version": "0.1.0",
+  "embedding_dim": 384,
+  "chemistry": "dna_r10.4.1_e8.2_400bps",
+  "molecule_type": "dna",
+  "artifacts": [
+    {
+      "file": "backbone.pt",
+      "size_bytes": 90747343,
+      "file_sha256": "4367c9a7ef291108a4bfcf6ae8452723c06bb68aa89f715ad4b8115b560986ac",
+      "role": "backbone",
+      "format": "pytorch_checkpoint",
+      "tensor_fingerprint": {
+        "scheme": "porefm-tensor-content-v1",
+        "sha256": "2528c42fb35af7c6d1e2c58e629e26567a28d469e2317f72f0977cf5214407de"
+      }
+    },
+    {
+      "file": "pretraining.pt",
+      "size_bytes": 115212771,
+      "file_sha256": "d120e0e9abba56fe8ec62cf8549e5fb969c84321982a0d55d4113fe670fdcf2f",
+      "role": "pretraining",
+      "format": "pytorch_checkpoint",
+      "tensor_fingerprint": {
+        "scheme": "porefm-tensor-content-v1",
+        "sha256": "4af4a3ef8b0cd0d4a0285f7fe316d7c53d230414f94e38b9ca1a773e4e1a967a"
+      }
+    }
+  ],
+  "tensor_fingerprint_scheme": {
+    "name": "porefm-tensor-content-v1",
+    "algorithm": "sha256",
+    "scope": "Stored tensors and numeric arrays, including their names, shapes and dtypes; scalar configuration and other metadata are excluded.",
+    "encoding": [
+      "Initialize SHA-256 with ASCII porefm-tensor-content-v1 followed by one zero byte.",
+      "Recursively traverse dictionaries sorted by (type(key).__name__, str(key)), and lists/tuples in index order. Skip leaves other than torch.Tensor and numpy.ndarray.",
+      "For each leaf, compact ASCII-escaped UTF-8 JSON encodes [key_path, dtype, shape, nbytes]. key_path is the list of dictionary keys and sequence indices. dtype is torch: plus str(tensor.dtype), or numpy: plus array.dtype.str.",
+      "Append the JSON byte length as an unsigned 8-byte big-endian integer, the JSON bytes, then C-contiguous numerical bytes in little-endian representation. Tensor/array storage padding is excluded."
+    ]
+  }
+}

pretraining.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d120e0e9abba56fe8ec62cf8549e5fb969c84321982a0d55d4113fe670fdcf2f
+size 115212771