Lev 9B: frozen weights (rounds 1-6)

NOTICE

@@ -1,4 +1,4 @@
-d3-mini (Decision 3.0)
+d3-flash (Decision 3.0)
 Copyright 2026 vLLM Semantic Router Team
 
-Built on Qwen/Qwen3.5-4B (Apache-2.0).
+Built on Qwen/Qwen3.5-9B (Apache-2.0).

config.json

@@ -15,9 +15,9 @@
     "full_attention_interval": 4,
     "head_dim": 256,
     "hidden_act": "silu",
-    "hidden_size": 2560,
+    "hidden_size": 4096,
     "initializer_range": 0.02,
-    "intermediate_size": 9216,
+    "intermediate_size": 12288,
     "layer_types": [
       "linear_attention",
       "linear_attention",
@@ -80,25 +80,25 @@
       "rope_theta": 10000000,
       "rope_type": "default"
     },
-    "tie_word_embeddings": true,
+    "tie_word_embeddings": false,
     "use_cache": true,
     "vocab_size": 248320
   },
-  "tie_word_embeddings": true,
+  "tie_word_embeddings": false,
   "transformers_version": "5.17.0",
   "video_token_id": 248057,
   "vision_config": {
     "deepstack_visual_indexes": [],
-    "depth": 24,
+    "depth": 27,
     "hidden_act": "gelu_pytorch_tanh",
-    "hidden_size": 1024,
+    "hidden_size": 1152,
     "in_channels": 3,
     "initializer_range": 0.02,
-    "intermediate_size": 4096,
+    "intermediate_size": 4304,
     "model_type": "qwen3_5_vision",
     "num_heads": 16,
     "num_position_embeddings": 2304,
-    "out_hidden_size": 2560,
+    "out_hidden_size": 4096,
     "patch_size": 16,
     "rope_parameters": {
       "rope_theta": 10000.0,

decision_config.json

@@ -2,8 +2,8 @@
   "format_version": 1,
   "format_id": "d3-code-readout-v1",
   "prompt": "d3",
-  "base_model": "Qwen/Qwen3.5-4B",
-  "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+  "base_model": "Qwen/Qwen3.5-9B",
+  "revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
   "codes": [
     "A",
     "B",
@@ -518,7 +518,7 @@
     12185,
     78382
   ],
-  "temperature": 1.25,
+  "temperature": 1.6,
   "attention_mode": "noncausal_full_attention",
   "pooling": "last",
   "max_length": null,

model-00001-of-00002.safetensors → model-00001-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:60621cb8c3f90e634b98f0ffaf609159ee2c450630386f09dd5f7fd7938d7598
-size 4973332144
+oid sha256:2c322016b91b1232c0d7628560164560d06d477a3edc3362714635f0e2f5949c
+size 4911663824

model-00002-of-00002.safetensors → model-00002-of-00004.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ab80e570e5387e617351b57585d0d9408aaa871e5e82e93d19a1ddcb4e978896
-size 4105283936
+oid sha256:169ae981a723d83037c340e0641bda9e2160c0383433dc8d9b274e236eb74b82
+size 4987839568

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:feb85fdf87b29d88a487fe5ee056a434a76dc55e09db0252fa8e80b7e946573a
+size 4954726768

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37ece9f15f7e4a7cf933b4201741172af73668ad5da8f0503919dbafa3bae3fe
+size 1931248816

questions.py

@@ -201,3 +201,78 @@ QUESTIONS.update({
 CITATOR = ("treatment", "history", "reported", "dissent")
 DECISION_LEVEL = ("outcome", "decision_type", "proceeding", "area", "publication_ban", "self_represented", "separate_opinions")
 PASSAGE_LEVEL = ("standard_of_review", "element_finding")
+
+
+# ---- Second-generation questions ----
+
+VOICE = {
+    "own_reasoning": "The court cites or discusses the case in its own reasoning.",
+    "inside_quotation": "The case appears only inside a passage the court quotes from another source.",
+    "other_court_or_party": "The case appears only where the court reports another court's reasoning or a party's argument.",
+    "not_named": "The excerpt never actually names or cites this case.",
+}
+TREATMENT_GROUP = {
+    "positive": "The court follows the case or applies its rule.",
+    "neutral": "The court considers the case or merely refers to it.",
+    "distinguished": "The court says the case does not govern because the facts or the law differ.",
+    "negative": "The court questions the case, declines to follow it, overrules it, or holds it superseded by statute.",
+    "not_courts_own": "The case appears only in a quotation, in a party's argument or another court's reasoning, or not at all.",
+    "same_matter": "The case is an earlier decision in the same proceeding.",
+}
+GROUP_OF = {"followed": "positive", "applied": "positive", "considered": "neutral", "referred_to": "neutral",
+            "distinguished": "distinguished", "questioned": "negative", "not_followed": "negative", "overruled": "negative",
+            "superseded_by_statute": "negative", "inside_quotation": "not_courts_own", "other_court_or_party": "not_courts_own",
+            "not_named": "not_courts_own", "same_matter": "same_matter"}
+VOICE_OF = {**{k: "own_reasoning" for k in GROUP_OF}, "inside_quotation": "inside_quotation",
+            "other_court_or_party": "other_court_or_party", "not_named": "not_named"}
+BURDEN = {
+    "balance_of_probabilities": "The court applies the balance of probabilities to this issue.",
+    "beyond_reasonable_doubt": "The court applies proof beyond a reasonable doubt to this issue.",
+    "other": "The court states some other burden or standard of proof for this issue.",
+    "not_stated": "The excerpt states no burden or standard of proof for this issue.",
+}
+PROVISION = {
+    "interpreted": "The court states what the provision's words mean, how far it reaches, or what its purpose requires.",
+    "applied": "The court applies the provision to the facts without stating its meaning.",
+    "mentioned": "The court only refers to the provision or quotes it.",
+}
+PREVAILING = {
+    "appellant": None, "respondent": None, "applicant": None, "plaintiff": None, "defendant": None, "accused": None,
+    "crown": None, "split": "The result is mixed.", "other": "Another party, or unclear.",
+}
+COUNSEL = {
+    "listed": "Counsel are named.",
+    "unnamed_representative": "A party was represented but the representative is not named.",
+    "none": "The decision says a party was unrepresented and no counsel is named.",
+    "redacted": "Names of counsel were removed or reduced to initials.",
+    "not_stated": "There is no appearance information at all.",
+}
+
+QUESTIONS.update({
+    "voice": {"type": "choice", "instructions": "In whose voice does the cited case appear in the citing decision?", "criteria": VOICE},
+    "negative": {"type": "noul", "instructions": "Does the citing court itself treat the cited case negatively: question it, decline to follow it, overrule it, or hold it superseded by statute?"},
+    "treatment_group": {"type": "choice", "instructions": "Broadly, how does the citing decision treat the cited case? If it treats the case in more than one way, choose the most negative.", "criteria": TREATMENT_GROUP},
+    "burden": {"type": "choice", "instructions": "Which burden or standard of proof does the court apply to the issue stated in the excerpt?", "criteria": BURDEN},
+    "provision_treatment": {"type": "choice", "instructions": "How does the decision treat the statutory provision named in the excerpt?", "criteria": PROVISION},
+    "proposition": {"type": "noul", "instructions": "Does the court, in this paragraph, state a proposition of law that it relies on to decide?"},
+    "prevailing_party": {"type": "choice", "instructions": "Which party prevailed in this decision?", "criteria": PREVAILING},
+    "counsel_status": {"type": "choice", "instructions": "What does the decision say about who appeared for the parties?", "criteria": COUNSEL},
+})
+AUXILIARY = ("voice", "negative", "treatment_group")
+SECOND_GENERATION = ("burden", "provision_treatment", "proposition", "prevailing_party", "counsel_status", "cited_for", "cite_fill")
+
+# questions whose options come with the example
+ITEM_INSTRUCTIONS = {
+    "cite_fill": "One citation was removed from the paragraph and replaced with [CITATION]. Which case was cited there?",
+    "cited_for": "What does the citing decision cite the case for?",
+    "au_treatment": "How does the citing decision treat the cited case? If it treats the case in more than one way, choose the most negative treatment.",
+}
+
+
+def spec_of(row):
+    """The question of a training or test row: a fixed one from QUESTIONS, or one the row carries itself."""
+    if row.get("options"):
+        return {"type": "choice", "instructions": row.get("instructions") or ITEM_INSTRUCTIONS[row["q"]], "criteria": dict(row["options"])}
+    if row.get("instructions"):
+        return {"type": "noul", "instructions": row["instructions"]}
+    return QUESTIONS[row["q"]]

readout.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b3f46e7aded29c1282cfc51d34a8462a9289a179d29aa44c912637a183e3c678
-size 2611280
+oid sha256:c1fcd628c2891dd37d7987a9195f354950123aa461830ab0a66cda4f49739745
+size 4178000