Update config.json

config.json

@@ -10,10 +10,15 @@
   },
   "backbone_type": "qwen3_5",
   "base_model": "Qwen/Qwen3.5-4B",
-  "calibrated": false,
+  "calibrated": true,
   "calibration": {
-    "method": "none",
-    "temperature": 1.0
+    "method": "joint-scalar-temperature",
+    "temperature": 1.0026751894886363,
+    "fit_split": "held-out temperature-calibration split",
+    "fit_rows": 1448,
+    "fit_objective": "hard_label_mean_nll",
+    "report_sha256": "d2fabc5aaf4012aceb8a3e8d0f1308060d9ed1876133c12ce575547bf7c33c3f",
+    "source_checkpoint": "private training checkpoint ensemble"
   },
   "code_scheme": "letters-v1",
   "codes": [
@@ -276,40 +281,26 @@
   "decision_protocol": "lemondecider-v1",
   "dtype": "bfloat16",
   "image_token_id": 248056,
-  "initialization_step": 0,
-  "max_input_length": 262144,
-  "max_options": 255,
-  "merge_provenance": {
-    "created_utc": "2026-10-15T07:10:31Z",
-    "derivation": "complete 0.5*A + 0.5*B FP32 tensor blend of two self-trained routes from a common Qwen3.5-4B initialization; FP32 readout island retained; code/tokenizer/protocol bundle from route B; v2 renames only machine identifiers, weights byte-identical to the v1 rebranded export",
-    "new_optimizer_steps": 0,
-    "new_training_performed": false,
-    "parents": [
-      {
-        "decision_protocol": "autojev-v1",
-        "lane": "route1",
-        "model_id": "route1-sft-ckpt2446-20261013",
-        "readout_dtype_source": "bfloat16",
-        "step": 2446
-      },
-      {
-        "decision_protocol": "lemondecider-v1",
-        "lane": "route2",
-        "model_id": "route2-sft-ckpt700-20261013",
-        "readout_dtype_source": "float32",
-        "step": 700
-      }
-    ],
-    "schema": "lemonjev-full-weight-merge/2"
+  "initialization_step": 4690,
+  "joint_readout": {
+    "marker": "<|fim_suffix|>",
+    "max_length": null,
+    "max_questions": null,
+    "readout_offset": 0,
+    "template": "joint-question-inline-readout-v1",
+    "usage": "nonpadding_joint_sequence_including_template_and_image_positions",
+    "version": 3
   },
-  "model_id": "lemondecider-4b-merge50-20261015-v2",
+  "max_input_length": 8192,
+  "max_options": 255,
+  "model_id": "lemondecider-4b-20261023-v3",
   "model_type": "lemondecider_decision",
   "pooling": "last",
-  "project_name": "LemonJev",
+  "project_name": "LemonDecider",
   "runtime_override": false,
-  "runtime_version": "lemonjev-qwen-decision-runtime-v2",
-  "source_model_id": "lemondecider-4b-merge50-20261015-v2",
-  "temperature": 1.0,
+  "runtime_version": "lemondecider-qwen-decision-runtime-v3",
+  "source_model_id": "lemondecider-v3-training-source",
+  "temperature": 1.0026751894886363,
   "text_config": {
     "attention_bias": false,
     "attention_dropout": 0.0,
@@ -650,8 +641,6 @@
   ],
   "trained": true,
   "transformers_version": "5.17.0",
-  "v2_change": "internal machine-identifier spelling normalized (lemondecider); weights byte-identical to v1",
-  "v2_of": "lemondecider-4b-merge50-20261015-v1",
   "video_token_id": 248057,
   "vision_config": {
     "deepstack_visual_indexes": [],
@@ -676,5 +665,18 @@
   },
   "vision_end_token_id": 248054,
   "vision_start_token_id": 248053,
-  "vision_training_mode": "frozen"
-}
+  "vision_training_mode": "frozen",
+  "lineage": {
+    "init": "Qwen3.5-4B initialization (bidirectional full-attention, FP32 255-way readout)",
+    "corpus": "1,199,692 rows: pplx-v1.1 625k + pplx-v1.2 DI train splits/tasksource 378k + startlux-v1.1 196k",
+    "recipe": "1 epoch joint-question-inline-readout-v1, lr 2e-6, EBS 256, 8k cap, fp32-master-bf16",
+    "soup": "uniform fp32 mean of four selected training checkpoints",
+    "calibration": {
+      "method": "joint-scalar-temperature",
+      "temperature": 1.0026751894886363,
+      "fit_rows": 1448
+    }
+  },
+  "author": "忍把浮名，换了浅斟低唱",
+  "author_line": "Author：忍把浮名，换了浅斟低唱"
+}