Upload README.md with huggingface_hub

README.md

@@ -14,16 +14,16 @@ tags:
 - decision-making
 ---
 
-# ⚡ Tasksource-JEV-Nano-v1 (`tasksource-jev-nano-v1`)
+# ⚡ Tasksource-JEV-Nano (`tasksource-jev-nano-v1`)
 
 **SOTA High-Throughput Native Decision Model (\<200M Parameters)**
 *Decoupled Multi-Vector Late Interaction for Fast, Calibrated System-One Decisions*
 
-[![Decision Index](https://img.shields.io/badge/Decision_Index_Raw-32.03-blue)](https://github.com/sileod/decision-models) [![Decision Index Skill](https://img.shields.io/badge/Decision_Index_Skill-10.32-green)](https://github.com/sileod/decision-models)
+[![Decision Index](https://img.shields.io/badge/Decision_Index_Raw-32.03-blue)](https://github.com/sileod/decision-models) [![Decision Index Skill](https://img.shields.io/badge/Decision_Index_Skill-10.32-green)](https://github.com/sileod/decision-models) [![Parameters](https://img.shields.io/badge/Parameters-149M-orange.svg)](https://huggingface.co/answerdotai/ModernBERT-base) [![Context Length](https://img.shields.io/badge/Context-8192-purple.svg)](https://huggingface.co/answerdotai/ModernBERT-base)
 
-Successor to [`tasksource-jev-nano-v0`](https://huggingface.co/tasksource/tasksource-jev-nano-v0) (Raw 30.65 / Skill 9.58, kept immutable for reproducibility). v1 is a single continued checkpoint — no soup — trained with plain grouped multi-question exposure (16k unique requests) on top of the v0 lineage. It uses the simple release inference path: state + single question (`S + Q_i`) with independently encoded options.
+Single 149M-parameter checkpoint (no soup): plain grouped multi-question exposure on the v0 lineage. Previous weights scored Raw 30.65 / Skill 9.58; current weights score Raw 32.03 / Skill 10.32 on the same full official suite.
 
-Built on [`answerdotai/ModernBERT-base`](https://huggingface.co/answerdotai/ModernBERT-base) and [`lightonai/LateOn`](https://huggingface.co/lightonai/LateOn) (~149M params, 8192 context). Handles choice (K=2 to 150+), NOUL, and score decisions.
+Built on [`answerdotai/ModernBERT-base`](https://huggingface.co/answerdotai/ModernBERT-base) and [`lightonai/LateOn`](https://huggingface.co/lightonai/LateOn). Handles choice (K=2 to 150+), NOUL, and score decisions.
 
 ## 📊 Benchmark Evaluations
 
@@ -42,36 +42,56 @@ Built on [`answerdotai/ModernBERT-base`](https://huggingface.co/answerdotai/Mode
 - **Tools & Automation**: 24.92% (Skill: 13.50%)
 - **Knowledge & Reasoning**: 23.63% (Skill: 2.12%)
 
+### 🏆 Sub-200M Peer Comparison
+
+| Model | Size | Decision Index Raw | Decision Index Skill |
+| --- | --- | --- | --- |
+| **Tasksource-JEV-Nano-v1** | **149M** | **32.03** 🏆 | **10.32** 🏆 |
+| Tasksource-JEV-Nano-v0 | 149M | 30.65 | 9.58 |
+| GLiClass Modern-Large | 340M | — | — |
+| GLiNER2.5 Base | 200M | — | — |
+
 ## 🚀 Usage
 
-```python
-from modernjev.multivector.lateon import LateOnDecisionScorer
-
-scorer = LateOnDecisionScorer(model_name_or_path="tasksource/tasksource-jev-nano-v1")
-out = scorer.decide_uncached(
-    states=["Customer reports unauthorized international wire transfer of $4,500."],
-    questions=["Select the appropriate fraud mitigation routing:"],
-    options=[["approve and monitor silently", "challenge with a push notification",
-              "freeze the account and call the customer", "decline and file a report"]],
-    kinds=["choice"],
-)
-print(out.probabilities)
+```bash
+pip install -U pylate torch
 ```
 
-Or raw PyLate:
-
 ```python
 from pylate import models
+import torch
+
 model = models.ColBERT(model_name_or_path="tasksource/tasksource-jev-nano-v1")
-ctx = model.encode(["<state>\nQuestion: <question>"], is_query=False, convert_to_tensor=True)[0]
-opts = model.encode(["<option 1>", "<option 2>"], is_query=True, convert_to_tensor=True)
+
+# 1. Decision context (state + question), encoded once
+state = "Customer reports unauthorized international wire transfer of $4,500."
+question = "Select the appropriate fraud mitigation routing:"
+context = [state + "\nQuestion: " + question]
+
+# 2. Candidate options, encoded independently
+options = [
+    "approve and monitor silently",
+    "challenge with a push notification",
+    "freeze the account and call the customer",
+    "decline and file a report",
+]
+
+# 3. Token-level MaxSim scores, temperature softmax
+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 = 0.3199  # learned model temperature
+probs = torch.softmax(scores / temperature, dim=0)
+
+for opt, p in zip(options, probs):
+    print(f"  {opt:45s} -> {p:.4f}")
 ```
 
 ## 📜 Citation
 
 ```bibtex
-@misc{sileo2026jevnano,
-  title={Tasksource-JEV-Nano: Decoupled Multi-Vector Late Interaction for Typed Decisions},
+@misc{sileo2026jevnanov1,
+  title={Tasksource-JEV-Nano-v1: Decoupled Multi-Vector Late Interaction for Typed Decisions},
   author={Sileo, Damien},
   year={2026},
   howpublished={\url{https://huggingface.co/tasksource/tasksource-jev-nano-v1}},