Update jeba-en adapter, choice head (r=32), and per-language temperature (v0.2.0)

README.md

@@ -62,19 +62,22 @@ split (`training/fit_calibration.py`). Configs and seed live under `training/con
 Reported by `training/evaluate.py` and rendered by `benchmarks/report.py` (accuracy, ECE, p50/p95
 latency per primitive and language).
 
-**Full-scale run (single RTX 3060 12GB):** 6,000 train / 1,500 eval deterministic synthetic
-records (localized cue terms per language, a learnable `other` team with rich descriptions, and
-distractor clauses), LoRA (r=16) plus a dedicated low-rank `choice` head (near-identity init),
-3 epochs, batch 16, bf16 + gradient checkpointing.
+**Full-scale run (single RTX 3060 12GB):** 9,000 English / 18,000 multilingual train / 1,500 eval
+deterministic synthetic records (fully localized per language, a learnable `other` team with rich
+descriptions, per-record RNG, one-in-six distractor clauses), LoRA (r=16) plus a dedicated
+low-rank `choice` head (r=32, near-identity init), 4 epochs, batch 16, bf16 + gradient
+checkpointing.
 
 | Checkpoint | Accuracy | ECE (calibrated) | p50 (ms) |
 | --- | --- | --- | --- |
-| English (ModernBERT-large + LoRA + choice head) | 0.721 | 0.052 | 18.4 |
-| Multilingual (mmBERT-base + LoRA + choice head) | 0.609 | 0.086 | 12.6 |
-
-Per primitive (English): `choice` 0.782, `noul` 0.718, `score` 0.664; (multilingual): `noul`
-0.728, `score` 0.700, `choice` 0.398. The dedicated `choice` head lifted English `choice` from
-~0.25 (chance) to 0.78; multilingual `choice` and `score` calibration remain the next targets.
+| English (ModernBERT-large + LoRA + choice head) | 0.781 | 0.077 | 22.9 |
+| Multilingual (mmBERT-base + LoRA + choice head) | 0.711 | 0.073 | 12.4 |
+
+Per primitive (English): `choice` 0.708, `noul` 0.744, `score` 0.892; (multilingual): `choice`
+0.734, `noul` 0.716, `score` 0.682. The localized, per-record-RNG data (B-1) lifted multilingual
+`choice` from 0.40 to 0.73 and English overall from 0.72 to 0.78. Per-language calibration for a
+few multilingual languages (`es`, `nl`, `de`) and multilingual `score` remain the next targets.
+The CUDA-graph fast path (`JEBA_FAST=1`) gives a 2.7× p50 speedup with 0 top-label flips.
 Full tables and environment are in
 [`benchmarks/report.md`](https://github.com/munod/jeba/blob/main/benchmarks/report.md).
 

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf711f809e17f1495c3849eeeb97189b466baacf52b0fb80e735d5726027549a
+oid sha256:a34daab0b0a3294f64a1eb921b5939a1bbd0a66e59344d6ca064c2684b14f22b
 size 28813352

finetune_config.json

@@ -1,21 +1,21 @@
 {
-  "batch_size": 16,
+  "batch_size": 8,
   "bf16": true,
   "choice_init_std": 0.01,
-  "choice_rank": 16,
+  "choice_rank": 32,
   "data_path": "data/train_en.jsonl",
-  "epochs": 3,
+  "epochs": 4,
   "grad_accum": 4,
   "gradient_checkpointing": true,
   "learning_rate": 0.0001,
   "lora_alpha": 32,
   "lora_dropout": 0.05,
   "lora_rank": 16,
-  "max_len": 256,
+  "max_len": 512,
   "max_records": null,
   "model_id": "answerdotai/ModernBERT-large",
   "out_dir": "checkpoints/en",
   "seed": 42,
   "use_4bit": false,
   "val_split": 0.1
-}
+}

temperature_calibration.json

@@ -2,22 +2,46 @@
   "bins": 10,
   "per_primitive": {
     "choice": {
-      "ece_after": 0.037886,
-      "ece_before": 0.037886,
+      "by_language": {
+        "en": {
+          "ece_after": 0.140568,
+          "ece_before": 0.191749,
+          "n": 500,
+          "temperature": 0.75
+        }
+      },
+      "ece_after": 0.140568,
+      "ece_before": 0.191749,
       "n": 500,
-      "temperature": 1.0
+      "temperature": 0.75
     },
     "noul": {
-      "ece_after": 0.027909,
-      "ece_before": 0.027909,
+      "by_language": {
+        "en": {
+          "ece_after": 0.011864,
+          "ece_before": 0.258169,
+          "n": 500,
+          "temperature": 0.75
+        }
+      },
+      "ece_after": 0.011864,
+      "ece_before": 0.258169,
       "n": 500,
-      "temperature": 1.0
+      "temperature": 0.75
     },
     "score": {
-      "ece_after": 0.148855,
-      "ece_before": 0.279323,
+      "by_language": {
+        "en": {
+          "ece_after": 0.089698,
+          "ece_before": 0.415709,
+          "n": 500,
+          "temperature": 0.05
+        }
+      },
+      "ece_after": 0.089698,
+      "ece_before": 0.415709,
       "n": 500,
-      "temperature": 0.25
+      "temperature": 0.05
     }
   },
   "warnings": []