Upload fine-tuned nullspan 10-op System One extraction model (Apple Silicon MPS)

README.md

@@ -25,7 +25,7 @@ pipeline_tag: question-answering
 ## Architectural Enhancements (`NullspanForExtraction`)
 
 Built on `distilbert/distilbert-base-uncased` (67M parameters) with 5 parallel System One heads and 10 routed operations:
-1. **`null_head`**: Calibrated `p_null` probability (`ECE: 1e-05`, `Brier: 0.0`) for absent field abstention (`value: null` + nearest `guess` decoy + `abstention_reason`).
+1. **`null_head`**: Calibrated `p_null` probability (`ECE: 0.01409`, `Brier: 0.00666`) for absent field abstention (`value: null` + nearest `guess` decoy + `abstention_reason`).
 2. **`span_head`**: Token-level start/end pointer network mapped via `offset_mapping` to exact character spans `[start, end]` with calibrated posterior `confidence` and Shannon `span_entropy_bits`.
 3. **`bio_head`**: Class-weighted token-level `B/I/O` sequence tagger extracting ordered multi-span arrays (`type: "list"`).
 4. **`op_head`**: 10-way System One computation, normalization, boolean judgment (`BOOL_JUDGE`), and zero-shot enum classification (`ENUM_CLASSIFY`) router.
@@ -33,12 +33,12 @@ Built on `distilbert/distilbert-base-uncased` (67M parameters) with 5 parallel S
 
 ## Multi-Axis Empirical Evaluation (`gnumanth/nullspan-dataset`)
 
-- **Training time**: `358.3s` (`Apple Silicon MPS`)
-- **Null Abstention Accuracy (`p_null`)**: `100.0%` (`ECE: 1e-05`, `Brier: 0.0`)
-- **Grounded Span Exact Match (`[start, end]`)**: `99.31%`
+- **Training time**: `208.3s` (`Apple Silicon MPS`)
+- **Null Abstention Accuracy (`p_null`)**: `99.14%` (`ECE: 0.01409`, `Brier: 0.00666`)
+- **Grounded Span Exact Match (`[start, end]`)**: `98.85%`
 - **10-Way Operation Router Accuracy**: `100.0%`
 - **Multi-Span List BIO Token Accuracy**: `100.0%`
-- **Quantile Distribution MAE / Pinball Loss**: `0.0007` / `0.0001`
+- **Quantile Distribution MAE / Pinball Loss**: `0.0008` / `0.0001`
 - **Canonical Multi-Task Suite Accuracy**: `100.0%`
 - **Adversarial Decoy & Prompt-Injection Stress Accuracy**: `100.0%`
 - **Cross-Field Invariant Verification Accuracy**: `100.0%`
@@ -82,26 +82,26 @@ print(result["answers"])
 {
   "reporter": {
     "type": "text",
-    "p_null": 3.8e-07,
+    "p_null": 1.455e-05,
     "span": [
       276,
       280
     ],
     "guess": "Mira",
-    "confidence": 0.999992,
-    "span_entropy_bits": 0.0002,
+    "confidence": 0.999944,
+    "span_entropy_bits": 0.0007,
     "value": "Mira"
   },
   "incident_id": {
     "type": "text",
-    "p_null": 3e-08,
+    "p_null": 1.19e-06,
     "span": [
       9,
       18
     ],
     "guess": "INC-40928",
-    "confidence": 0.999962,
-    "span_entropy_bits": 0.0006,
+    "confidence": 0.999976,
+    "span_entropy_bits": 0.0004,
     "value": "INC-40928"
   },
   "opened_on": {
@@ -118,27 +118,27 @@ print(result["answers"])
   },
   "total_credit": {
     "type": "number",
-    "p_null": 4.5e-06,
+    "p_null": 6.73e-06,
     "span": [
       109,
       130
     ],
     "guess": 1550.75,
     "confidence": 0.999989,
-    "span_entropy_bits": 0.0002,
+    "span_entropy_bits": 0.0001,
     "value": 1550.75
   },
   "outage_days": {
     "type": "duration",
-    "p_null": 3.04e-06,
+    "p_null": 3.45e-06,
     "unit": "days",
     "span": [
       29,
       41
     ],
     "guess": 7,
-    "confidence": 0.999992,
-    "span_entropy_bits": 0.0001,
+    "confidence": 0.999986,
+    "span_entropy_bits": 0.0002,
     "value": 7
   },
   "failure_modes": {
@@ -157,12 +157,12 @@ print(result["answers"])
         100
       ]
     ],
-    "confidence": 0.497096,
-    "span_entropy_bits": 1.5381
+    "confidence": 0.999761,
+    "span_entropy_bits": 0.0034
   },
   "severity": {
     "type": "enum",
-    "p_null": 1.4e-07,
+    "p_null": 5.33e-06,
     "span": [
       155,
       160
@@ -173,38 +173,38 @@ print(result["answers"])
       "sev-3"
     ],
     "probs": {
-      "sev-1": 0.999975,
-      "sev-2": 1.3e-05,
-      "sev-3": 1.3e-05
+      "sev-1": 0.99999,
+      "sev-2": 5e-06,
+      "sev-3": 5e-06
     },
     "guess": "sev-1",
-    "confidence": 0.999975,
-    "span_entropy_bits": 0.0005,
+    "confidence": 0.99999,
+    "span_entropy_bits": 0.0001,
     "value": "sev-1"
   },
   "data_loss": {
     "type": "boolean",
-    "p_null": 1.28e-06,
-    "p_true": 7e-06,
+    "p_null": 3.4e-06,
+    "p_true": 8e-06,
     "span": [
       188,
       193
     ],
     "guess": false,
-    "confidence": 0.999993,
+    "confidence": 0.999992,
     "span_entropy_bits": 0.0001,
     "value": false
   },
   "cve_id": {
     "type": "text",
-    "p_null": 0.9999429,
+    "p_null": 0.99949825,
     "span": [
       123,
       130
     ],
     "guess": "$310.25",
     "confidence": 0.05,
-    "span_entropy_bits": 0.0009,
+    "span_entropy_bits": 0.0011,
     "value": null,
     "abstention_reason": "unassigned_placeholder"
   },
@@ -212,11 +212,11 @@ print(result["answers"])
     "type": "estimate",
     "value": 1890,
     "median": 1870,
-    "p10": 1270,
+    "p10": 1260,
     "p90": 2670,
     "p_below": 0.0,
     "p_above": 0.0,
-    "confidence": 0.99999
+    "confidence": 0.999931
   }
 }
 ```

eval_results.json

@@ -1,53 +1,53 @@
 {
   "repo_id": "gnumanth/nullspan",
   "device": "Apple Silicon MPS",
-  "train_seconds": 358.28,
-  "epochs": 6,
-  "train_samples": 2090,
-  "val_samples": 336,
+  "train_seconds": 208.29,
+  "epochs": 4,
+  "train_samples": 2330,
+  "val_samples": 348,
   "val_metrics": {
-    "null_accuracy": 100.0,
-    "null_ece": 1e-05,
-    "null_brier_score": 0.0,
-    "span_exact_match": 99.31,
+    "null_accuracy": 99.14,
+    "null_ece": 0.01409,
+    "null_brier_score": 0.00666,
+    "span_exact_match": 98.85,
     "op_accuracy": 100.0,
     "list_bio_token_accuracy": 100.0,
-    "quantile_mae_norm": 0.0007,
+    "quantile_mae_norm": 0.0008,
     "quantile_pinball_loss": 0.0001
   },
   "canonical_eval": {
     "canonical_field_accuracy": 100.0,
     "adversarial_stress_accuracy": 100.0,
     "invariant_verification_accuracy": 100.0,
-    "mean_doc_latency_ms": 472.43,
+    "mean_doc_latency_ms": 259.24,
     "examples": {
       "nullspan_incident_postmortem": {
-        "latency_ms": 86.15,
+        "latency_ms": 108.95,
         "num_fields": 10,
         "invariants_passed": true,
         "answers": {
           "reporter": {
             "type": "text",
-            "p_null": 3.8e-07,
+            "p_null": 1.455e-05,
             "span": [
               276,
               280
             ],
             "guess": "Mira",
-            "confidence": 0.999992,
-            "span_entropy_bits": 0.0002,
+            "confidence": 0.999944,
+            "span_entropy_bits": 0.0007,
             "value": "Mira"
           },
           "incident_id": {
             "type": "text",
-            "p_null": 3e-08,
+            "p_null": 1.19e-06,
             "span": [
               9,
               18
             ],
             "guess": "INC-40928",
-            "confidence": 0.999962,
-            "span_entropy_bits": 0.0006,
+            "confidence": 0.999976,
+            "span_entropy_bits": 0.0004,
             "value": "INC-40928"
           },
           "opened_on": {
@@ -64,27 +64,27 @@
           },
           "total_credit": {
             "type": "number",
-            "p_null": 4.5e-06,
+            "p_null": 6.73e-06,
             "span": [
               109,
               130
             ],
             "guess": 1550.75,
             "confidence": 0.999989,
-            "span_entropy_bits": 0.0002,
+            "span_entropy_bits": 0.0001,
             "value": 1550.75
           },
           "outage_days": {
             "type": "duration",
-            "p_null": 3.04e-06,
+            "p_null": 3.45e-06,
             "unit": "days",
             "span": [
               29,
               41
             ],
             "guess": 7,
-            "confidence": 0.999992,
-            "span_entropy_bits": 0.0001,
+            "confidence": 0.999986,
+            "span_entropy_bits": 0.0002,
             "value": 7
           },
           "failure_modes": {
@@ -103,12 +103,12 @@
                 100
               ]
             ],
-            "confidence": 0.497096,
-            "span_entropy_bits": 1.5381
+            "confidence": 0.999761,
+            "span_entropy_bits": 0.0034
           },
           "severity": {
             "type": "enum",
-            "p_null": 1.4e-07,
+            "p_null": 5.33e-06,
             "span": [
               155,
               160
@@ -119,38 +119,38 @@
               "sev-3"
             ],
             "probs": {
-              "sev-1": 0.999975,
-              "sev-2": 1.3e-05,
-              "sev-3": 1.3e-05
+              "sev-1": 0.99999,
+              "sev-2": 5e-06,
+              "sev-3": 5e-06
             },
             "guess": "sev-1",
-            "confidence": 0.999975,
-            "span_entropy_bits": 0.0005,
+            "confidence": 0.99999,
+            "span_entropy_bits": 0.0001,
             "value": "sev-1"
           },
           "data_loss": {
             "type": "boolean",
-            "p_null": 1.28e-06,
-            "p_true": 7e-06,
+            "p_null": 3.4e-06,
+            "p_true": 8e-06,
             "span": [
               188,
               193
             ],
             "guess": false,
-            "confidence": 0.999993,
+            "confidence": 0.999992,
             "span_entropy_bits": 0.0001,
             "value": false
           },
           "cve_id": {
             "type": "text",
-            "p_null": 0.9999429,
+            "p_null": 0.99949825,
             "span": [
               123,
               130
             ],
             "guess": "$310.25",
             "confidence": 0.05,
-            "span_entropy_bits": 0.0009,
+            "span_entropy_bits": 0.0011,
             "value": null,
             "abstention_reason": "unassigned_placeholder"
           },
@@ -158,133 +158,133 @@
             "type": "estimate",
             "value": 1890,
             "median": 1870,
-            "p10": 1270,
+            "p10": 1260,
             "p90": 2670,
             "p_below": 0.0,
             "p_above": 0.0,
-            "confidence": 0.99999
+            "confidence": 0.999931
           }
         }
       },
       "nullspan_gpu_invoice": {
-        "latency_ms": 924.51,
+        "latency_ms": 192.3,
         "num_fields": 9,
         "invariants_passed": true,
         "answers": {
           "provider": {
             "type": "text",
-            "p_null": 1.79e-06,
+            "p_null": 1.003e-05,
             "span": [
               13,
               29
             ],
             "guess": "TensorCore Cloud",
-            "confidence": 0.99998,
-            "span_entropy_bits": 0.0003,
+            "confidence": 0.999874,
+            "span_entropy_bits": 0.0019,
             "value": "TensorCore Cloud"
           },
           "invoice_number": {
             "type": "number",
-            "p_null": 2.067e-05,
+            "p_null": 6.792e-05,
             "span": [
               39,
               44
             ],
             "guess": 88310,
-            "confidence": 0.999978,
-            "span_entropy_bits": 0.0,
+            "confidence": 0.999925,
+            "span_entropy_bits": 0.0002,
             "value": 88310
           },
           "issued_at": {
             "type": "datetime",
-            "p_null": 2.04e-06,
+            "p_null": 2.14e-06,
             "span": [
               55,
               71
             ],
             "guess": "2026-08-14T18:20:00",
-            "confidence": 0.999998,
+            "confidence": 0.999996,
             "span_entropy_bits": 0.0,
             "value": "2026-08-14T18:20:00"
           },
           "nodes": {
             "type": "number",
-            "p_null": 9.7e-07,
+            "p_null": 2.6e-06,
             "span": [
               79,
               80
             ],
             "guess": 4,
-            "confidence": 0.999998,
+            "confidence": 0.999997,
             "span_entropy_bits": 0.0,
             "value": 4
           },
           "hourly_rate": {
             "type": "number",
-            "p_null": 7.08e-06,
+            "p_null": 1.619e-05,
             "span": [
               84,
               88
             ],
             "guess": 3.5,
-            "confidence": 0.999989,
+            "confidence": 0.999981,
             "span_entropy_bits": 0.0001,
             "value": 3.5
           },
           "subtotal": {
             "type": "number",
-            "p_null": 1.948e-05,
+            "p_null": 5.004e-05,
             "span": [
               105,
               110
             ],
             "guess": 14,
-            "confidence": 0.999979,
-            "span_entropy_bits": 0.0,
+            "confidence": 0.999947,
+            "span_entropy_bits": 0.0001,
             "value": 14,
             "verified": true
           },
           "egress_fee": {
             "type": "number",
-            "p_null": 1.42e-06,
+            "p_null": 3.92e-06,
             "span": [
               124,
               128
             ],
             "guess": 0,
-            "confidence": 0.999995,
-            "span_entropy_bits": 0.0001,
+            "confidence": 0.999981,
+            "span_entropy_bits": 0.0003,
             "value": 0
           },
           "priority_surcharge": {
             "type": "number",
-            "p_null": 6.9e-07,
+            "p_null": 2.43e-06,
             "span": [
               150,
               154
             ],
             "guess": 1.75,
-            "confidence": 0.999998,
+            "confidence": 0.999997,
             "span_entropy_bits": 0.0,
             "value": 1.75
           },
           "promo_rebate": {
             "type": "number",
-            "p_null": 0.99995482,
+            "p_null": 0.99978632,
             "span": [
               124,
               128
             ],
             "guess": 0,
             "confidence": 0.05,
-            "span_entropy_bits": 0.0001,
+            "span_entropy_bits": 0.0003,
             "value": null,
             "abstention_reason": "missing_field"
           }
         }
       },
       "nullspan_cluster_comps": {
-        "latency_ms": 513.59,
+        "latency_ms": 706.33,
         "num_fields": 1,
         "invariants_passed": true,
         "answers": {
@@ -292,148 +292,148 @@
             "type": "estimate",
             "value": 23200,
             "median": 23000,
-            "p10": 20400,
+            "p10": 20300,
             "p90": 26100,
             "p_below": 0.0,
             "p_above": 0.0,
-            "confidence": 0.579559
+            "confidence": 0.987942
           }
         }
       },
       "nullspan_edge_forecast": {
-        "latency_ms": 365.46,
+        "latency_ms": 29.38,
         "num_fields": 1,
         "invariants_passed": true,
         "answers": {
           "requests": {
             "type": "forecast",
             "value": [
-              2416041
+              2417871
             ],
             "median": [
-              2418555
+              2420386
             ],
             "p10": [
-              2263103
+              2265913
             ],
             "p90": [
-              2568073
+              2559063
             ],
-            "confidence": 0.055277
+            "confidence": 0.999989
           }
         }
       }
     },
     "adversarial_examples": {
       "adv_prompt_injection_and_decoy_id": {
-        "latency_ms": 1400.38,
+        "latency_ms": 111.04,
         "num_fields": 4,
         "invariants_passed": true,
         "answers": {
           "incident_id": {
             "type": "text",
-            "p_null": 3e-08,
+            "p_null": 1.29e-06,
             "span": [
               9,
               18
             ],
             "guess": "INC-77401",
-            "confidence": 0.999991,
+            "confidence": 0.999988,
             "span_entropy_bits": 0.0002,
             "value": "INC-77401"
           },
           "cve_id": {
             "type": "text",
-            "p_null": 0.99984634,
+            "p_null": 0.9986155,
             "span": [
               116,
               123
             ],
             "guess": "$180.00",
             "confidence": 0.05,
-            "span_entropy_bits": 0.003,
+            "span_entropy_bits": 0.0017,
             "value": null,
             "abstention_reason": "unassigned_placeholder"
           },
           "total_credit": {
             "type": "number",
-            "p_null": 1.567e-05,
+            "p_null": 1.739e-05,
             "span": [
               103,
               123
             ],
             "guess": 1080,
-            "confidence": 0.999972,
-            "span_entropy_bits": 0.0003,
+            "confidence": 0.999953,
+            "span_entropy_bits": 0.0005,
             "value": 1080
           },
           "data_loss": {
             "type": "boolean",
-            "p_null": 9.7e-07,
-            "p_true": 6e-06,
+            "p_null": 3.05e-06,
+            "p_true": 1.4e-05,
             "span": [
               181,
               186
             ],
             "guess": false,
-            "confidence": 0.999994,
-            "span_entropy_bits": 0.0001,
+            "confidence": 0.999986,
+            "span_entropy_bits": 0.0002,
             "value": false
           }
         }
       },
       "adv_negated_rebate_and_invariant": {
-        "latency_ms": 135.73,
+        "latency_ms": 60.62,
         "num_fields": 4,
         "invariants_passed": true,
         "answers": {
           "nodes": {
             "type": "number",
-            "p_null": 9.3e-07,
+            "p_null": 2.54e-06,
             "span": [
               77,
               78
             ],
             "guess": 6,
-            "confidence": 0.999998,
+            "confidence": 0.999997,
             "span_entropy_bits": 0.0,
             "value": 6
           },
           "hourly_rate": {
             "type": "number",
-            "p_null": 1.581e-05,
+            "p_null": 3.146e-05,
             "span": [
               82,
               87
             ],
             "guess": 12.5,
-            "confidence": 0.999983,
+            "confidence": 0.999967,
             "span_entropy_bits": 0.0001,
             "value": 12.5
           },
           "subtotal": {
             "type": "number",
-            "p_null": 2.108e-05,
+            "p_null": 5.559e-05,
             "span": [
               104,
               109
             ],
             "guess": 75,
-            "confidence": 0.999977,
+            "confidence": 0.999941,
             "span_entropy_bits": 0.0001,
             "value": 75,
             "verified": true
           },
           "promo_rebate": {
             "type": "number",
-            "p_null": 0.9999516,
+            "p_null": 0.99977404,
             "span": [
               123,
               127
             ],
             "guess": 0,
             "confidence": 0.05,
-            "span_entropy_bits": 0.0003,
+            "span_entropy_bits": 0.0008,
             "value": null,
             "abstention_reason": "missing_field"
           }
@@ -444,75 +444,51 @@
   "history": [
     {
       "epoch": 1,
-      "train_loss": 3.7146,
-      "null_accuracy": 100.0,
-      "null_ece": 0.0103,
-      "null_brier_score": 0.00245,
-      "span_exact_match": 99.31,
+      "train_loss": 2.5958,
+      "null_accuracy": 97.13,
+      "null_ece": 0.02155,
+      "null_brier_score": 0.01361,
+      "span_exact_match": 96.84,
       "op_accuracy": 100.0,
-      "list_bio_token_accuracy": 99.84,
-      "quantile_mae_norm": 0.0015,
-      "quantile_pinball_loss": 0.0003
+      "list_bio_token_accuracy": 100.0,
+      "quantile_mae_norm": 0.0014,
+      "quantile_pinball_loss": 0.0002
     },
     {
       "epoch": 2,
-      "train_loss": 0.1044,
-      "null_accuracy": 100.0,
-      "null_ece": 2e-05,
-      "null_brier_score": 0.0,
-      "span_exact_match": 99.31,
+      "train_loss": 0.5978,
+      "null_accuracy": 97.41,
+      "null_ece": 0.0197,
+      "null_brier_score": 0.0113,
+      "span_exact_match": 97.13,
       "op_accuracy": 100.0,
       "list_bio_token_accuracy": 100.0,
-      "quantile_mae_norm": 0.0012,
+      "quantile_mae_norm": 0.0013,
       "quantile_pinball_loss": 0.0002
     },
     {
       "epoch": 3,
-      "train_loss": 0.0433,
-      "null_accuracy": 100.0,
-      "null_ece": 1e-05,
-      "null_brier_score": 0.0,
-      "span_exact_match": 99.31,
+      "train_loss": 0.3431,
+      "null_accuracy": 97.99,
+      "null_ece": 0.0191,
+      "null_brier_score": 0.01109,
+      "span_exact_match": 98.28,
       "op_accuracy": 100.0,
       "list_bio_token_accuracy": 100.0,
       "quantile_mae_norm": 0.0011,
-      "quantile_pinball_loss": 0.0002
+      "quantile_pinball_loss": 0.0001
     },
     {
       "epoch": 4,
-      "train_loss": 0.027,
-      "null_accuracy": 100.0,
-      "null_ece": 1e-05,
-      "null_brier_score": 0.0,
-      "span_exact_match": 99.31,
-      "op_accuracy": 100.0,
-      "list_bio_token_accuracy": 100.0,
-      "quantile_mae_norm": 0.0008,
-      "quantile_pinball_loss": 0.0002
-    },
-    {
-      "epoch": 5,
-      "train_loss": 0.0223,
-      "null_accuracy": 100.0,
-      "null_ece": 1e-05,
-      "null_brier_score": 0.0,
-      "span_exact_match": 99.31,
+      "train_loss": 0.2589,
+      "null_accuracy": 99.14,
+      "null_ece": 0.01409,
+      "null_brier_score": 0.00666,
+      "span_exact_match": 98.85,
       "op_accuracy": 100.0,
       "list_bio_token_accuracy": 100.0,
       "quantile_mae_norm": 0.0008,
       "quantile_pinball_loss": 0.0001
-    },
-    {
-      "epoch": 6,
-      "train_loss": 0.0176,
-      "null_accuracy": 100.0,
-      "null_ece": 1e-05,
-      "null_brier_score": 0.0,
-      "span_exact_match": 99.31,
-      "op_accuracy": 100.0,
-      "list_bio_token_accuracy": 100.0,
-      "quantile_mae_norm": 0.0007,
-      "quantile_pinball_loss": 0.0001
     }
   ]
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:096fbde74dcff6a07fefea60ed56a4496f3d831d4946e9e1d6e44d1c5dd0c1a1
+oid sha256:2b8c5342c51d7f2c7f59b3245743486c159ab5ffe408b72330c32ae423b98e8e
 size 268258852

modeling_nullspan.py

@@ -66,6 +66,41 @@ def format_field_query(field_name: str, spec: dict[str, Any]) -> str:
     return " | ".join(parts)
 
 
+_QA_STOPWORDS = {
+    "what", "who", "when", "where", "which", "whom", "whose", "why", "how",
+    "was", "were", "did", "does", "do", "is", "are", "been", "being", "be",
+    "the", "a", "an", "in", "on", "of", "to", "for", "with", "from", "by",
+    "at", "into", "unto", "upon", "about", "as", "that", "this", "these",
+    "those", "their", "his", "her", "its", "they", "them", "it", "he", "she",
+    "had", "has", "have", "would", "could", "should", "can", "may", "might",
+}
+
+
+def select_context_window(state: str, spec: dict[str, Any]) -> tuple[str, int]:
+    """Select a focused 380-char prose window around the highest-overlap sentence when state > 380 chars."""
+    if len(state) <= 380 or "\n" in state or spec.get("type", "text") != "text":
+        return state, 0
+    instr = str(spec.get("instructions", ""))
+    q_toks = [t.lower() for t in re.sub(r"[^\w\s-]", " ", instr).split() if len(t) > 1 and t.lower() not in _QA_STOPWORDS]
+    if not q_toks:
+        return state, 0
+    sentences = list(re.finditer(r"[^.!?\n]+[.!?\n]?", state))
+    best_m = None
+    best_sc = -1.0
+    for m_s in sentences:
+        s_low = m_s.group(0).lower()
+        hits = sum(1.0 for t in q_toks if t in s_low or t.rstrip("s") in s_low)
+        if hits > best_sc:
+            best_sc = hits
+            best_m = m_s
+    if best_m is None or best_m.end() <= 340:
+        return state, 0
+    win_start = max(0, best_m.start() - 40)
+    while win_start > 0 and state[win_start - 1] not in (" ", "\n"):
+        win_start -= 1
+    return state[win_start:], win_start
+
+
 def extract_numeric_context_features(state: str, spec: dict[str, Any]) -> list[float]:
     """Extract 8 normalized numeric summary features from `state` for the quantile head."""
     vmin = float(spec.get("min", 1.0))
@@ -315,13 +350,8 @@ class NullspanForExtraction(PreTrainedModel):
 
         if is_null_labels is not None:
             loss_null = F.binary_cross_entropy_with_logits(null_logits, is_null_labels.float())
-            valid_sp = start_positions != -100
-            if valid_sp.any():
-                loss_start = F.cross_entropy(start_logits[valid_sp], start_positions[valid_sp])
-                loss_end = F.cross_entropy(end_logits[valid_sp], end_positions[valid_sp])
-            else:
-                loss_start = null_logits.sum() * 0.0
-                loss_end = null_logits.sum() * 0.0
+            loss_start = F.cross_entropy(start_logits, start_positions.clamp(min=0))
+            loss_end = F.cross_entropy(end_logits, end_positions.clamp(min=0))
 
             valid_bio = bio_labels != -100
             if valid_bio.any():
@@ -361,14 +391,16 @@ class NullspanForExtraction(PreTrainedModel):
         state: str,
         fields: dict[str, dict[str, Any]],
         tokenizer: Any,
-        max_length: int = 384,
+        max_length: int = 128,
     ) -> dict[str, Any]:
         """Run a single batched System One forward pass across all requested fields."""
         self.eval()
         device = next(self.parameters()).device
         field_names = list(fields.keys())
         queries = [format_field_query(fn, fields[fn]) for fn in field_names]
-        states = [state for _ in field_names]
+        win_pairs = [select_context_window(state, fields[fn]) for fn in field_names]
+        states = [wp[0] for wp in win_pairs]
+        win_offsets = [wp[1] for wp in win_pairs]
         num_feats_list = [extract_numeric_context_features(state, fields[fn]) for fn in field_names]
 
         enc = tokenizer(
@@ -432,7 +464,8 @@ class NullspanForExtraction(PreTrainedModel):
                             best_score = sc
                             best_s, best_e = s_i, e_i
 
-            char_s, char_e = int(offsets[best_s][0]), int(offsets[best_e][1])
+            w_off = win_offsets[b_i]
+            char_s, char_e = int(offsets[best_s][0]) + w_off, int(offsets[best_e][1]) + w_off
             span_text = state[char_s:char_e].strip(" .,;:\n")
             if span_text:
                 adj_idx = state.find(span_text, char_s)

onnx/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71ee96c7db97c911af0145a171e0615b789fa57f9455b1edce8a927bd0a7fae3
+size 268368196

onnx/model_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70315e1726bf2aa631c4be42e5a054d3fdc699b4b22b71c8affdda663be5f2d8
+size 67476041

tokenizer.json

@@ -2,7 +2,7 @@
   "version": "1.0",
   "truncation": {
     "direction": "Right",
-    "max_length": 384,
+    "max_length": 128,
     "strategy": "OnlySecond",
     "stride": 0
   },