Decision 2.0 package DEV2.0-2B (release) from 78592cbb436eb3c963a6c6927110481a15db53c0-src_training_decision2

MODEL_MANIFEST.json

@@ -2,7 +2,7 @@
   "base": null,
   "builder": {
     "module_sha256": "cbb19d06cfb1c6db4d6513de33da4eaff15e6dc933e726fa1475f3c91305667d",
-    "source_commit": "0a50a86b802b8aab4918ad6211a9736602484a17"
+    "source_commit": "78592cbb436eb3c963a6c6927110481a15db53c0"
   },
   "calibration": null,
   "card": {
@@ -24,9 +24,9 @@
     "assets/jevarena-v3-rank.svg": "26a36aa0569d9176357be3438320145c07cedbcc280d65864828c435b77ea17b",
     "assets/jevbench-public231-rank.svg": "7ed4bf852ef67c95c91ad8a85ba7faba53bdde0115f6a30d492dad2a7fb7ed4e",
     "backbone/config.json": "071e97d8291168ba712237744c9a60e733acce554e6164322d22550dc96de19b",
-    "backbone/model-00001-of-00002.safetensors": "1f083a66d8cdcd02887452b2801efae6af41a6e9529bdc37c962b994721dca08",
-    "backbone/model-00002-of-00002.safetensors": "3a3f291be8d6ac079f1f737ec89092ed3947abd2cd5fd342823c91ce1084af8b",
-    "backbone/model.safetensors.index.json": "12b12d9183d20062b4b5ab51d0d0a666ada0a2c53b8bd1b078f3456d5c3db781",
+    "backbone/model-00001-of-00002.safetensors": "17ba386b5f9bab7c648ef2509988165cdc341bb75c35a2fa5e8b27faece88fe0",
+    "backbone/model-00002-of-00002.safetensors": "aec9a7eaadce4e7ad764614a56a8183d67ee94eb2e63969165ba129d07a0281a",
+    "backbone/model.safetensors.index.json": "7607bb56aa11afe3a32bba1855e6bc3e54767b4a29c28216765f368b01da45a2",
     "chat_template.jinja": "273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80",
     "config.json": "3d08526e2defacc65e64bfa6fe575d77bed7b320513edac99000ca5507a3e20a",
     "decision2/__init__.py": "10fea99a87686a8f5d99e9a8a91573ce1ddb7bc702f22b1fa21d3de258a4054a",
@@ -50,16 +50,16 @@
   "identity": {
     "fingerprint_files": {
       "backbone/config.json": "071e97d8291168ba712237744c9a60e733acce554e6164322d22550dc96de19b",
-      "backbone/model-00001-of-00002.safetensors": "1f083a66d8cdcd02887452b2801efae6af41a6e9529bdc37c962b994721dca08",
-      "backbone/model-00002-of-00002.safetensors": "3a3f291be8d6ac079f1f737ec89092ed3947abd2cd5fd342823c91ce1084af8b",
-      "backbone/model.safetensors.index.json": "12b12d9183d20062b4b5ab51d0d0a666ada0a2c53b8bd1b078f3456d5c3db781",
+      "backbone/model-00001-of-00002.safetensors": "17ba386b5f9bab7c648ef2509988165cdc341bb75c35a2fa5e8b27faece88fe0",
+      "backbone/model-00002-of-00002.safetensors": "aec9a7eaadce4e7ad764614a56a8183d67ee94eb2e63969165ba129d07a0281a",
+      "backbone/model.safetensors.index.json": "7607bb56aa11afe3a32bba1855e6bc3e54767b4a29c28216765f368b01da45a2",
       "chat_template.jinja": "273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80",
       "decision_config.json": "0b6c3429ee06032739d7386659c332ed7bb62d0b96ccddfcad4fad999e22cd4f",
       "decision_head.safetensors": "33b6541bb6636677eb91a4d8e06acd4db81152b11097088840796f11d49c707a",
       "tokenizer.json": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
       "tokenizer_config.json": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87"
     },
-    "model_sha256": "073bd1f2fe62e39fe993f57006bab17ece50a7e6fc7c5ee72107fefddeb81da4"
+    "model_sha256": "32872f2968e38aa99901797aaab5e12e32e281bf67bb924037d443394411170d"
   },
   "kind": "release",
   "licence": {
@@ -125,7 +125,7 @@
   "profile": "qwen-full",
   "repo_id": "llm-semantic-router/DEV2.0-2B",
   "runtime": {
-    "equivalence": "decision2/qwen.py loads this full checkpoint with the vendored training/model sources whose SHA-256 equal the scored adapter sources (checked at build time) and applies the per-item batching, BF16-backbone / FP32-head execution, raw probabilities (temperature 1; no calibration file) and answer normalization of v2.dec.infer_dec, which for this non-residual checkpoint wraps the same DecisionModel without extra readouts. Checked on one GPU of the scoring node against the T = 1 predictions derived exactly from the sealed CAL698 predictions of every scored prompt (typed-final 1,600, css15 6,547, public231 231) and of the mlx-diag diagnostic (2,275) by release.sh --parity, with the scored run's persisted Triton autotune cache.",
+    "equivalence": "decision2/qwen.py loads this full checkpoint with the vendored training/model sources whose SHA-256 equal the scored adapter sources (checked at build time) and applies the per-item batching, BF16-backbone / FP32-head execution, raw probabilities (temperature 1; no calibration file) and answer normalization of v2.dec.infer_dec, which for this non-residual checkpoint wraps the same DecisionModel without extra readouts. The scored checkpoint (073bd1f2) stored every tensor in FP32; this package (v2.release.bf16_copy, receipt a5229ef1) stores its 186 Linear projection matrices in BF16 exactly as BF16 autocast rounds them and every other tensor bit for bit in FP32. Checked on one GPU of the scoring node against the T = 1 predictions derived exactly from the sealed CAL698 predictions of every scored prompt (typed-final 1,600, css15 6,547, public231 231) and of the mlx-diag diagnostic (2,275) by release.sh --parity, with the scored run's persisted Triton autotune cache.",
     "requirements": {
       "causal-conv1d": "1.7.0 (GPU convolution kernels)",
       "flash-linear-attention": "0.5.2 (GPU gated-delta kernels)",

backbone/model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f083a66d8cdcd02887452b2801efae6af41a6e9529bdc37c962b994721dca08
-size 3999855496
+oid sha256:17ba386b5f9bab7c648ef2509988165cdc341bb75c35a2fa5e8b27faece88fe0
+size 3017470920

backbone/model-00002-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3a3f291be8d6ac079f1f737ec89092ed3947abd2cd5fd342823c91ce1084af8b
-size 3527479552
+oid sha256:aec9a7eaadce4e7ad764614a56a8183d67ee94eb2e63969165ba129d07a0281a
+size 1764429584

backbone/model.safetensors.index.json

@@ -1,7 +1,7 @@
 {
   "metadata": {
     "total_parameters": 1881825088,
-    "total_size": 7527300352
+    "total_size": 4781866240
   },
   "weight_map": {
     "embed_tokens.weight": "model-00001-of-00002.safetensors",