Decision 2.0 package Decision-2.0-Sol-2B (release) from 088818218df33b6c2dca39badffde47962125d41-src_training_decision2

MODEL_MANIFEST.json

@@ -2,7 +2,7 @@
   "base": null,
   "builder": {
     "module_sha256": "7f74839fb11762d66fa4cdc9657e7dfa8f4041bb20293dedbf69d54640fe0f0d",
-    "source_commit": "807acc299469e887b63a794830d3e547786a132f"
+    "source_commit": "088818218df33b6c2dca39badffde47962125d41"
   },
   "calibration": null,
   "card": {
@@ -15,11 +15,11 @@
       "assets/jevarena.png": "0d97e2ecf10414cf211a37e8f8f16fbebddbc1dc5a85f4a243b1ce029d489081"
     },
     "index_sha256": "97a95897fb13b81addf033dee7984d53e996ddb8baad5076f706a0b7c752e22e",
-    "readme_sha256": "7b91af2a48bf47e9ef67c774a98720e062948cf4f8c5c8957a6a3c491b9ec360"
+    "readme_sha256": "0c547477c690faabd9e066c0fa51fa98e9d5388e225bea4703e3e5a9dfe08c24"
   },
   "files_sha256": {
     "LICENSE": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
-    "README.md": "7b91af2a48bf47e9ef67c774a98720e062948cf4f8c5c8957a6a3c491b9ec360",
+    "README.md": "0c547477c690faabd9e066c0fa51fa98e9d5388e225bea4703e3e5a9dfe08c24",
     "assets/banner.png": "2363db1a11ea3d8c61656a078c2b2cdf10623177e54ca3ac5a48b22ccb0be724",
     "assets/index-areas.png": "3670b08f2c186c03a78028ad0c0c9ab07a4475ffb33b81ce88daea600b1a0d37",
     "assets/index-pareto.png": "bbc3b1c70755ff97ed5943bfa65f5b1c1e24041b0923c102610104182db5ecba",
@@ -71,12 +71,12 @@
       {
         "component": "Decision-2.0-Sol-2B weights, decision head, package runtime, card and artwork",
         "licence": "apache-2.0",
-        "source": "llm-semantic-router/Decision-2.0-Sol-2B"
+        "source": "vllm-sr/Decision-2.0-Sol-2B"
       },
       {
         "component": "Decision 1.0 Sol-2B weights (direct weight origin, fully fine-tuned)",
         "licence": "apache-2.0",
-        "source": "llm-semantic-router/Decision-1.0-Sol-2B@ce0c018a"
+        "source": "vllm-sr/Decision-1.0-Sol-2B@ce0c018a"
       },
       {
         "component": "Qwen3.5-2B text backbone and tokenizer (upstream of Decision 1.0 Sol)",
@@ -101,7 +101,7 @@
   "model_name": "Decision-2.0-Sol-2B",
   "origin": {
     "relation": "finetune",
-    "repo_id": "llm-semantic-router/Decision-1.0-Sol-2B",
+    "repo_id": "vllm-sr/Decision-1.0-Sol-2B",
     "revision": "ce0c018a28de16d6639b1cd203b761bf643b89e6",
     "summary": "Every weight of Decision 1.0 Sol was fine-tuned (nothing frozen, no adapter). The release is the uniform average of two seeds of that fine-tune, trained with the previous Decision 2.0 Sol 2B's answer probabilities as soft targets (self-distillation) and with extra copies of released multilingual rows that keep the released multilingual token share. Decision 1.0 Sol is itself a text-only fine-tune of [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) at `15852e8c16360a2fea060d615a32b45270f8a8fc` (Apache-2.0), whose text backbone and tokenizer this model inherits; the Qwen3.5 vision tower is not part of it."
   },
@@ -172,7 +172,7 @@
       "5.18.0"
     ]
   },
-  "repo_id": "llm-semantic-router/Decision-2.0-Sol-2B",
+  "repo_id": "vllm-sr/Decision-2.0-Sol-2B",
   "runtime": {
     "equivalence": "decision2/qwen.py loads this full checkpoint with the training/model sources vendored from the scored run's own runner mirror (5b246b110), 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 (8c8e98e3) stored every tensor in FP32; this package (v2.release.bf16_copy) stores its Linear projection matrices in BF16 exactly as BF16 autocast rounds them and every other tensor bit for bit in FP32, and the runtime holds the backbone's BF16-exact Linear weights in BF16, the values BF16 autocast multiplies with. The runtime, the Transformers remote code (AutoModel with trust_remote_code) and the forward token budget are the current revision's. Checked on one GPU, in the scored image with a copy of the persisted Triton autotune cache of the formal run, against the sealed formal predictions of every scored prompt (typed-final 1,600, css15 6,547, public231 231) by release.sh --parity before and after the download, and AutoModel against the native runtime on every scored prompt.",
     "requirements": {

README.md

@@ -1,6 +1,6 @@
 ---
 license: apache-2.0
-base_model: llm-semantic-router/Decision-1.0-Sol-2B
+base_model: vllm-sr/Decision-1.0-Sol-2B
 base_model_relation: finetune
 library_name: transformers
 tags:
@@ -14,7 +14,7 @@ tags:
 
 # Decision-2.0-Sol-2B
 
-**Decision-2.0-Sol-2B** is the 2B model of [Decision 2.0](https://huggingface.co/collections/llm-semantic-router/decision-20-6ab7cf7bdfb506bf8269cb00), the decision models of [vLLM Semantic Router](https://github.com/vllm-project/semantic-router). Give it an input (text or JSON) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all at once and returns a probability for every answer, without generating text.
+**Decision-2.0-Sol-2B** is the 2B model of [Decision 2.0](https://huggingface.co/collections/vllm-sr/decision-20-6ab7cf7bdfb506bf8269cb00), the decision models of [vLLM Semantic Router](https://github.com/vllm-project/semantic-router). Give it an input (text or JSON) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all at once and returns a probability for every answer, without generating text.
 
 | | |
 | --- | --- |
@@ -41,7 +41,7 @@ import json
 
 from transformers import AutoModel
 
-model = AutoModel.from_pretrained("llm-semantic-router/Decision-2.0-Sol-2B", trust_remote_code=True)
+model = AutoModel.from_pretrained("vllm-sr/Decision-2.0-Sol-2B", trust_remote_code=True)
 result = model.system_one(
     state="The order arrived damaged yesterday. The customer has a receipt and asks for a replacement today.",
     questions={
@@ -72,7 +72,7 @@ result = model.system_one(
 print(json.dumps(result["answers"], indent=2))
 
 # Or as a pipeline:
-# transformers.pipeline("decision", model="llm-semantic-router/Decision-2.0-Sol-2B", trust_remote_code=True)(state=..., questions=...)
+# transformers.pipeline("decision", model="vllm-sr/Decision-2.0-Sol-2B", trust_remote_code=True)(state=..., questions=...)
 ```
 
 ## Evaluation
@@ -112,6 +112,6 @@ Apache-2.0 ([LICENSE](LICENSE)).
   title        = {{Decision-2.0-Sol-2B}: A Decision 2.0 Model for Structured Decisions},
   author       = {{vLLM Semantic Router Team}},
   year         = {2026},
-  howpublished = {\url{https://huggingface.co/llm-semantic-router/Decision-2.0-Sol-2B}}
+  howpublished = {\url{https://huggingface.co/vllm-sr/Decision-2.0-Sol-2B}}
 }
 ```