Upload folder using huggingface_hub

README.md

@@ -27,8 +27,26 @@ pipeline_tag: text-classification
 > [`munod/tachyone-multi`](https://huggingface.co/munod/tachyone-multi)); measured numbers below come
 > from `benchmarks/report.md`.
 >
-> **This revision is B-15: both checkpoints retrained on `template_split: "train"`, and the
-> holdout number is honest for the first time.** The data recipe was corrected end to end
+> **This revision is B-17: the teacher expanded the tone banks, the multilingual arm was
+> retrained, and B-1's per-language ECE acceptance closes.** **352 phrases** authored by a
+> local `qwen3.6:35b` teacher went into every `neutral`/`request` bank (five domains × six
+> languages, inserted **before** the held-out phrase — the holdout evals are byte-identical,
+> sha-pinned); the multilingual checkpoint retrained with the unchanged B-15 recipe and
+> **English was not touched** (byte-identical numbers). On phrasing never seen in any
+> training: **multilingual 0.9264 (routed), English 0.9825** — the in-template → holdout gap
+> collapsed **−0.1105 → −0.0165** — and **all 12 `(primitive, language)` cells now sit at
+> ECE ≤ 0.05** (worst `score/fr` 0.0257; `score/it` **0.0521 → 0.0014**, accuracy
+> 0.8699 → 0.9928) — **B-1 criterion 3 (NFR-C06) CLOSED** on the frozen yardstick. Also
+> improved: `choice` holdout 0.9248 → 0.9808, the `noul` tripwire 0.7807 → 0.9735, every
+> language ≥ 0.9928 (`es` 1.0000), in-template multilingual ECE 0.0228 → 0.0036 (zero gate
+> exceptions everywhere). Disclosures, never smoothed: the expansion targeted a diagnosed
+> mechanism (BACKLOG **B-17** carries the per-domain table); off-domain **MASSIVE
+> 0.074 → 0.058 with `ECE raw` 0.046 → 0.197** — in-domain gains re-sharpened off-domain
+> confidence (the L-007/L-015 trade), still 3.4× its chance; and the English P3 `choice`
+> ECE **0.5243** (pooled fit at T = 10.0) stands exactly as previously disclosed.
+>
+> **The v0.9.0/v0.10.0 revisions were B-15/B-16 (history).** B-15 retrained both checkpoints
+> on `template_split: "train"` after the data recipe was corrected end to end
 > (sampler strides decoupled so every language trains every label, evals made
 > leave-one-template-out, phrase content ×3, volume 1.7×) and every training set regenerated so
 > training and the holdout eval share **0%** of templates. The multilingual lineage carries the
@@ -142,9 +160,9 @@ interleaved touch-up (`training/interleave_continue.py`).
 | English (ModernBERT-large + five-domain LoRA r=16 + choice-head bank) | support | **1.0000** | 0.188 | 57.7 |
 | English, five-domain split | 5 domains | **0.9983** | 0.191 | 57.7 |
 | English, **holdout phrasing** (never in training) | 5 domains | **0.9825** | 0.188 | 58.0 |
-| Multilingual (mmBERT-base + LoRA r=64 + fitted bank) — routed runtime | support | **0.9107** | 0.051 | 47.4 |
-| Multilingual, five-domain split | 5 domains | **0.9156** | 0.019 | 48.0 |
-| Multilingual, **holdout phrasing** (never in training) | 5 domains | **0.8355** | 0.103 | 47.8 |
+| Multilingual (mmBERT-base + LoRA r=64 + fitted bank) — routed runtime | support | **0.9300** | 0.038 | 52.8 |
+| Multilingual, five-domain split | 5 domains | **0.9221** | 0.026 | 53.5 |
+| Multilingual, **holdout phrasing** (never in training) | 5 domains | **0.9264** | 0.032 | 53.3 |
 
 The English ECE column reads the **pooled never-trained fit** (below): `choice` is flattened to
 T=10.0 to calibrate held-out rows, which shows up honestly in-domain (0.5243 `choice` ECE —
@@ -214,8 +232,11 @@ rows fall through to the English checkpoint by language routing, `.specs/project
 0.05 ECE target on that split, and on the **holdout** the B-1 per-language ECE criterion was
 **NOT met** at B-15 (7 cells, worst `score/it` 0.1495) — the **B-16** cycle (serve-key
 detector + fitted confidence, 2026-10-09) moved it to **11/12 with `score/it` 0.0521
-recorded NOT met**, `choice` 0.9248 and every other axis byte-identical (recorded with
-numbers in BACKLOG **B-1**/**B-16**) — open under NFR-C06 / BACKLOG B-1.
+recorded NOT met** — and the **B-17** cycle (teacher bank expansion + retrain, 2026-10-10)
+**closed it**: **all 12/12 cells ≤ 0.05 on the same frozen holdout** (worst `score/fr`
+0.0257, `score/it` **0.0014** at accuracy 0.9928), `choice` 0.9248 → 0.9808 and every other
+axis improved or byte-identical. **NFR-C06 met**; the trail is in BACKLOG
+**B-1**/**B-15**/**B-16**/**B-17**.
 The CUDA-graph fast path
 (`TACHYONE_FAST=1`) gives **2.53×** p50 on English (17.31 → 6.83 ms) and **3.71×** on the
 multilingual five-domain path (14.23 → 3.84 ms) — re-measured on the v0.9.0 artifacts

adapter_config.json

@@ -35,9 +35,9 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "Wo",
     "Wqkv",
-    "Wi"
+    "Wi",
+    "Wo"
   ],
   "target_parameters": null,
   "task_type": null,

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f7969603c1b3a443d413cc98ffdc423f0d35d759b439e6282784cdeaccdfd95d
+oid sha256:eb946aab6a74a3e5cbf1811c1c4cf85579941d1b56e9c9208da188848852b616
 size 54088096

choice_bank_fit.json

@@ -1,10 +1,10 @@
 {
   "arm": "bank",
   "config": {
-    "bank_dir": "checkpoints/multi_tt_il_fit_bank",
+    "bank_dir": "checkpoints/multi_b17_il_fit_bank",
     "batch_size": 64,
     "choice_rank": 128,
-    "control_dir": "checkpoints/multi_tt_il_fit_ctrl",
+    "control_dir": "checkpoints/multi_b17_il_fit_ctrl",
     "data_path": "data/train_multi_domains.jsonl",
     "device": "auto",
     "epochs": 8,
@@ -13,7 +13,7 @@
     "max_records": null,
     "model_id": "jhu-clsp/mmBERT-base",
     "seed": 42,
-    "trunk_adapter": "checkpoints/multi_tt_il",
+    "trunk_adapter": "checkpoints/multi_b17_il",
     "val_split": 0.1
   },
   "domains": [
@@ -26,10 +26,10 @@
   "fit": {
     "arm": "bank",
     "epochs": 8,
-    "temperature": 0.996136486530304,
-    "train_loss": 0.0011428351117203106,
+    "temperature": 0.9909090399742126,
+    "train_loss": 0.013521413194443138,
     "train_records": 15660,
-    "val_loss": 0.0011507968665770243,
+    "val_loss": 0.020827124596054517,
     "val_records": 1740
   },
   "languages": [
@@ -40,5 +40,5 @@
     "nl",
     "pt"
   ],
-  "trunk_adapter": "checkpoints/multi_tt_il"
+  "trunk_adapter": "checkpoints/multi_b17_il"
 }

choice_head.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:34409891d55ba3c09feac69d694c7e6326855699f869f8f90e9692737592acd4
-size 39934565
+oid sha256:2c7222ea220d28da546c067b10ab1bf343d5ff26d0c7fe1db08984ea9f21ac76
+size 39897981

temperature_calibration.json

@@ -4,134 +4,134 @@
     "choice": {
       "by_language": {
         "de": {
-          "ece_after": 0.008217,
-          "ece_before": 0.015348,
+          "ece_after": 0.013708,
+          "ece_before": 0.043495,
           "n": 830,
-          "temperature": 3.0
+          "temperature": 6.0
         },
         "es": {
-          "ece_after": 0.034263,
-          "ece_before": 0.073255,
+          "ece_after": 0.0,
+          "ece_before": 5e-06,
           "n": 840,
-          "temperature": 6.0
+          "temperature": 0.05
         },
         "fr": {
-          "ece_after": 0.003171,
-          "ece_before": 0.003911,
+          "ece_after": 0.002616,
+          "ece_before": 0.0096,
           "n": 830,
-          "temperature": 2.0
+          "temperature": 5.0
         },
         "it": {
-          "ece_after": 0.027815,
-          "ece_before": 0.063041,
+          "ece_after": 0.00251,
+          "ece_before": 0.00355,
           "n": 830,
-          "temperature": 6.0
+          "temperature": 3.0
         },
         "nl": {
-          "ece_after": 0.004955,
-          "ece_before": 0.010624,
+          "ece_after": 0.00271,
+          "ece_before": 0.005693,
           "n": 830,
-          "temperature": 3.0
+          "temperature": 4.0
         },
         "pt": {
-          "ece_after": 0.044754,
-          "ece_before": 0.076013,
+          "ece_after": 0.002972,
+          "ece_before": 0.006214,
           "n": 840,
-          "temperature": 6.0
+          "temperature": 5.0
         }
       },
-      "ece_after": 0.011135,
-      "ece_before": 0.040317,
+      "ece_after": 0.00348,
+      "ece_before": 0.011391,
       "n": 5000,
-      "temperature": 6.0
+      "temperature": 5.0
     },
     "noul": {
       "by_language": {
         "de": {
-          "ece_after": 0.022377,
-          "ece_before": 0.242584,
+          "ece_after": 0.001729,
+          "ece_before": 0.301935,
           "n": 830,
-          "temperature": 0.25
+          "temperature": 0.1
         },
         "es": {
-          "ece_after": 0.08133,
-          "ece_before": 0.211321,
+          "ece_after": 8.7e-05,
+          "ece_before": 0.303815,
           "n": 840,
-          "temperature": 0.25
+          "temperature": 0.05
         },
         "fr": {
-          "ece_after": 0.043647,
-          "ece_before": 0.26635,
+          "ece_after": 0.002239,
+          "ece_before": 0.299088,
           "n": 830,
-          "temperature": 0.15
+          "temperature": 0.1
         },
         "it": {
-          "ece_after": 0.027566,
-          "ece_before": 0.189535,
+          "ece_after": 0.005682,
+          "ece_before": 0.302041,
           "n": 830,
-          "temperature": 0.5
+          "temperature": 0.1
         },
         "nl": {
-          "ece_after": 0.046929,
-          "ece_before": 0.215839,
+          "ece_after": 0.006508,
+          "ece_before": 0.292156,
           "n": 830,
-          "temperature": 0.25
+          "temperature": 0.15
         },
         "pt": {
-          "ece_after": 0.060943,
-          "ece_before": 0.207737,
+          "ece_after": 0.000725,
+          "ece_before": 0.3015,
           "n": 840,
-          "temperature": 0.25
+          "temperature": 0.1
         }
       },
-      "ece_after": 0.053913,
-      "ece_before": 0.215547,
+      "ece_after": 0.00323,
+      "ece_before": 0.298892,
       "n": 5000,
-      "temperature": 0.25
+      "temperature": 0.15
     },
     "score": {
       "by_language": {
         "de": {
-          "ece_after": 0.018422,
-          "ece_before": 0.42241,
+          "ece_after": 0.001757,
+          "ece_before": 0.469981,
           "n": 830,
-          "temperature": 0.25
+          "temperature": 0.15
         },
         "es": {
-          "ece_after": 0.041522,
-          "ece_before": 0.40273,
+          "ece_after": 0.002654,
+          "ece_before": 0.444646,
           "n": 840,
           "temperature": 0.25
         },
         "fr": {
-          "ece_after": 0.012019,
-          "ece_before": 0.432135,
+          "ece_after": 0.002539,
+          "ece_before": 0.44168,
           "n": 830,
           "temperature": 0.25
         },
         "it": {
-          "ece_after": 0.049433,
-          "ece_before": 0.393083,
+          "ece_after": 0.004361,
+          "ece_before": 0.466521,
           "n": 830,
-          "temperature": 0.25
+          "temperature": 0.15
         },
         "nl": {
-          "ece_after": 0.021803,
-          "ece_before": 0.42233,
+          "ece_after": 0.003243,
+          "ece_before": 0.466974,
           "n": 830,
-          "temperature": 0.25
+          "temperature": 0.15
         },
         "pt": {
-          "ece_after": 0.011156,
-          "ece_before": 0.435601,
+          "ece_after": 0.009933,
+          "ece_before": 0.457518,
           "n": 840,
-          "temperature": 0.25
+          "temperature": 0.15
         }
       },
-      "ece_after": 0.024324,
-      "ece_before": 0.415722,
+      "ece_after": 0.0108,
+      "ece_before": 0.457823,
       "n": 5000,
-      "temperature": 0.25
+      "temperature": 0.15
     }
   },
   "warnings": []