v3 weights: noulhalved-1M + 10% procedural (cbv2 60.5/60.0, b77 47.8)

1_Dense/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:304a390aa5da090c9c22a999ab8fc18194172e97f2098b7bbff9dd1f68d6fa91
+oid sha256:7df0bcb1fea1070eb7a0d2e52204191cbb65ec7f83b6a4b5042b8c1f36149365
 size 9437360

2_Dense/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ca0a3107481100bbec203c119b39afa6d47345a31dcdc015a763ab1dbd8a256e
+oid sha256:bca58efc434526f8b3f1ae88d72a952cdcc44dad5de658b8bf0f5937792ef943
 size 9437360

3_Dense/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad5e8919ffbd769684d58e7e4a5219c14b2834e04dc0101a28a5f65d8eed5e14
+oid sha256:d996ba145030ba46d351f7e6f573bc161e1925f683a39bc6e780f7711bacddf9
 size 393304

README.md

@@ -14,37 +14,39 @@ language:
 base_model: lightonai/LateOn
 ---
 
-# Tasksource-JEV-Nano-v0
+# Tasksource-JEV-Nano-v3
 
 A small decision model that picks the best option from a list — fraud routings, intents, topics, sentiments — using token-level late interaction instead of a classifier head.
 
-It is LateOn (~149M parameters) fine-tuned on 512k real decisions covering three judgment types: picking one option (*choice*), yes/no questions (*noul*), and graded scores (*score*). No teacher model was used.
+It is LateOn (~149M parameters) fine-tuned on ~1.04M decisions covering three judgment types: picking one option (*choice*), yes/no questions (*noul*), and graded scores (*score*). No teacher model was used. **These v3 weights replace the original v0 weights** at this repo id.
 
 **Why this architecture.** The situation (state + question) is encoded once into token vectors and reused for every candidate set, while each option is encoded independently. That gives two properties classifier heads don't have: caching the situation across decisions, and outputs that don't depend on option order (verified exactly permutation-equivariant).
 
 - **Base model:** [`lightonai/LateOn`](https://huggingface.co/lightonai/LateOn)
-- **Training data:** [`tasksource/tasksource-jev-typed-decisions`](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions) (pinned revision [`071f0cf2201f06b2dea3e55323aa3e264284e139`](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions/tree/071f0cf2201f06b2dea3e55323aa3e264284e139)), 512k-decision stream, natural mix of the three judgment types, 519 tasks
-- **Checkpoint selection:** Tasksource unseen tasks only (step 4000 of 8000)
+- **Training data:** [`tasksource/tasksource-jev-typed-decisions`](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions) (pinned revision [`7470ed9e61ebd06fbec45e0c3239eb7a239bb1b4`](https://huggingface.co/datasets/tasksource/tasksource-jev-typed-decisions/tree/7470ed9e61ebd06fbec45e0c3239eb7a239bb1b4)), noul-halved natural stream (943,772 rows) + 100k procedural rows from [`tasksource/procedural-typed-decisions`](https://huggingface.co/datasets/tasksource/procedural-typed-decisions) (9.6% of the 1,043,772-decision stream, seed 13)
+- **Training:** 16,308 steps × 64, `sq_o` path, no distillation, no role tokens; checkpoint selection on Tasksource unseen-dev NLL (best step 12000, NLL 1.105)
 - **Training code:** `decision-models` repo (`scripts/train_multivector.py`, canonical run)
 
 ## Evaluation
 
-On Tasksource's held-out tasks across all three judgment types, v0 scores **46.8%** on unseen-dev and **49.7%** on unseen-test (1k examples each). It reaches **41.3% exact match** on Fast Decisions (2,900 cases across 17 domains). The table compares models on the same 1k examples per public dataset; classifier-benchmark v2 has 866 cases.
+Head-to-head under one eval harness (new-pin probes, 1k rows each; classifier-benchmark v2, 866 cases). v3 wins 5 of 7 suites, with a clear margin on Banking77 (+3.1) and classifier-benchmark v2 (+3.0 micro). Its one regression is typed decisions (−4.2): the procedural rows displace organic business-task coverage at fixed budget.
 
-### Compact peers (all under 200M)
+| Model | Unseen-dev | Unseen-test | Typed | AG News | Emotion | Banking77 | cb-v2 micro / macro |
+|---|---|---|---|---|---|---|---|
+| **v3 (this repo)** | **48.1** | 51.6 | 38.4 | **74.8** | **49.1** | **47.8** | **60.5 / 60.0** |
+| v0 (replaced) | 47.3 | **52.0** | **42.6** | 73.8 | 49.0 | 44.7 | 57.5 / 57.2 |
+
+### Compact peers (all under 200M, v0-era shared probes — reference only)
 
 | Model | Size | AG News | Emotion | Banking77 | Typed | cb-v2 micro |
 |---|---|---|---|---|---|---|
-| **Tasksource-JEV-Nano-v0** | 149M | 73.8 | 49.0 | 44.7 | 42.6 | 57.5 |
 | GLiClass Base v3 | 187M | 76.9 | 51.8 | 52.4 | 47.3 | 51.2 |
 | GLiClass Modern-Base v3 | 151M | 75.7 | 56.1 | 38.7 | 48.6 | 43.9 |
 | GLiNER2.5 Base | 194M | 76.2 | 56.6 | 69.7 | 43.9 | 60.0 |
 | GLiNER2.5 Small | 74M | 70.4 | 54.2 | 66.5 | 34.0 | 53.5 |
 | GLiClass Edge v3 | 33M | 65.5 | 50.1 | 23.2 | 40.0 | 35.3 |
 
-v0's classifier-benchmark v2 macro accuracy is **57.2%**.
-
-Larger peers (GLiClass Modern-Large 399M, GLiNER2.5-Multi 287M, Laya-Multilingual 322M) are omitted from this table; Laya's 90.0 AG News is training overlap (AG News is in its training mix). v0 trades some raw high-cardinality accuracy for state caching and order-independence, which none of the classifier heads offer.
+Larger peers (GLiClass Modern-Large 399M, GLiNER2.5-Multi 287M, Laya-Multilingual 322M) are omitted; Laya's 90.0 AG News is training overlap. v3 trades some raw high-cardinality accuracy for state caching and order-independence, which none of the classifier heads offer.
 
 ## Use it (only public packages)
 
@@ -73,7 +75,7 @@ ctx = model.encode(context, is_query=False, convert_to_tensor=True)[0]
 opts = model.encode(options, is_query=True, convert_to_tensor=True)
 scores = torch.stack([(o @ ctx.T).max(dim=1).values.sum() for o in opts])
 
-temperature = 0.53  # learned temperature of this release
+temperature = 0.354  # learned temperature of this release
 probs = torch.softmax(scores / temperature, dim=0)
 for opt, p in zip(options, probs):
     print(f"  {opt:45s} -> {p:.4f}")
@@ -81,11 +83,17 @@ for opt, p in zip(options, probs):
 
 Yes/no and graded decisions work the same way with two options or an ordered scale; encode the situation once and reuse it across as many candidate sets as you like.
 
+## Files
+
+- `model.safetensors` (596MB) + `1_Dense/`, `2_Dense/`, `3_Dense/` — fine-tuned LateOn backbone with the decision head
+- `decision_meta.json` — path mode (`sq_o`), temperature, token configuration needed to reproduce scoring
+- `config.json`, `tokenizer.json`, `modules.json` — standard sentence-transformers packaging
+
 ## Citation
 
 ```bibtex
-@misc{sileo2026jevnanov0,
-  title={Tasksource-JEV-Nano-v0: Decoupled Multi-Vector Late Interaction for Typed Decisions},
+@misc{sileo2026jevnanov3,
+  title={Tasksource-JEV-Nano-v3: Decoupled Multi-Vector Late Interaction for Typed Decisions},
   author={Sileo, Damien},
   year={2026},
   howpublished={\url{https://huggingface.co/tasksource/tasksource-jev-nano-v0}},

config.json

@@ -74,5 +74,5 @@
   "sparse_prediction": false,
   "tie_word_embeddings": true,
   "transformers_version": "5.17.0",
-  "vocab_size": 50370
+  "vocab_size": 50372
 }

decision_meta.json

@@ -1,8 +1,15 @@
 {
-  "init_temperature": 0.5347228646278381,
+  "init_temperature": 0.35398468375205994,
   "path_mode": "sq_o",
   "pooling_budget": null,
   "protected_tokens": 1,
+  "rel_slots": 2,
+  "joint_rel_distinct": false,
+  "rel_score_scale": 32.0,
+  "role_tokens": false,
+  "trimax_rank": null,
+  "has_query_gate": false,
+  "gate_dim": null,
   "config": {
     "model_name_or_path": "lightonai/LateOn",
     "embedding_dim": 128,
@@ -13,6 +20,11 @@
     "path_mode": "sq_o",
     "pooling_budget": null,
     "protected_tokens": 1,
+    "rel_slots": 2,
+    "joint_rel_distinct": false,
+    "rel_score_scale": null,
+    "role_tokens": false,
+    "trimax_rank": null,
     "query_prefix": "[Q] ",
     "document_prefix": "[D] ",
     "backend": "torch",

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:82b40602a9ac23a332640fce13e8f9b2ccf8703d9ac5d36b40f6d41107e2a92c
-size 596076280
+oid sha256:5c5a3a2a75c76ccef9fe1dc17cb2aa3b021f28c48d786f7221c2d06cd5e4fc5b
+size 596082424

tokenizer.json

@@ -1076,6 +1076,24 @@
       "rstrip": false,
       "normalized": true,
       "special": false
+    },
+    {
+      "id": 50370,
+      "content": "[REL0]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 50371,
+      "content": "[REL1]",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
     }
   ],
   "normalizer": {