Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
fix(ddp): fix calibration temperature array dimensions by avoiding 2D concatenation
Nandakishorfcf127c57176
diff --git a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynbindex 4837c1cf..808cbb7e 100644--- a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb+++ b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb@@ -349,8 +349,7 @@" torch.cuda.empty_cache()\n"," model.eval()\n"," calib_items = all_items[::15][:400]\n",- " \n",- " logits_cal_list = []\n",+ " calib_preds = []\n"," with torch.no_grad():\n"," for c_idx in range(0, len(calib_items), 16):\n"," c_chunk = calib_items[c_idx:c_idx + 16]\n",@@ -363,21 +362,20 @@" cb[\"marker_mask\"].to(device),\n"," cb[\"qtype\"].to(device)\n"," )\n",- " logits_cal_list.append(l_sub.float().cpu().numpy())\n",- " \n",- " logits_cal = np.concatenate(logits_cal_list, axis=0)\n",+ " l_np = l_sub.float().cpu().numpy()\n",+ " for r, it in enumerate(c_chunk):\n",+ " k = len(it[\"markers\"])\n",+ " calib_preds.append((it[\"qtype\"], l_np[r, :k], it[\"target\"]))\n"," \n"," fitted_temps = [1.2, 1.2, 1.2]\n"," try:\n"," for qt in range(3):\n",- " sel = [(logits_cal[idx, :len(calib_items[idx][\"markers\"])], calib_items[idx][\"target\"])\n",- " for idx in range(len(calib_items)) if calib_items[idx][\"qtype\"] == qt]\n",+ " sel = [(z, t) for q_type, z, t in calib_preds if q_type == qt]\n"," if sel:\n"," fitted_temps[qt] = fit_one_temp(sel)\n"," print(\"Fitted calibration temperatures (choice, score, noul):\", [round(t, 3) for t in fitted_temps])\n"," except Exception as e:\n"," print(\"Temperature fitting fallback:\", e)\n",- " \n"," os.makedirs(output_dir, exist_ok=True)\n"," sd = {k: v.half().contiguous().cpu() for k, v in model.state_dict().items()}\n"," save_file(sd, os.path.join(output_dir, \"model.safetensors\"))\n",