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
feat: increase token budget to 1024 tokens in Kaggle notebook
Nandakishora29dcf8cf181
diff --git a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynbindex c39021e6..6fc86180 100644--- a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb+++ b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb@@ -158,7 +158,7 @@"metadata": {},"source": ["%%writefile /kaggle/working/train_ddp.py\n",- "import os, sys, time, json, random\n",+ "import os, sys, time, json, random, math\n","import torch\n","import torch.distributed as dist\n","from torch.nn.parallel import DistributedDataParallel as DDP\n",@@ -191,6 +191,25 @@" \"label\": torch.tensor([it[\"label\"] for it in items])\n"," }\n","\n",+ "def fit_one_temp(sel):\n",+ " if len(sel) < 10:\n",+ " return 1.0\n",+ " kmax = max(len(z) for z, _ in sel)\n",+ " Z = torch.full((len(sel), kmax), -1e4)\n",+ " T = torch.zeros((len(sel), kmax))\n",+ " for i, (z, t) in enumerate(sel):\n",+ " Z[i, :len(z)] = torch.tensor(z)\n",+ " T[i, :len(t)] = torch.tensor(t, dtype=torch.float32)\n",+ " log_t = torch.zeros(1, requires_grad=True)\n",+ " opt = torch.optim.LBFGS([log_t], lr=0.1, max_iter=100)\n",+ " def closure():\n",+ " opt.zero_grad()\n",+ " loss = -(T * torch.log_softmax(Z / log_t.exp(), -1)).sum(-1).mean()\n",+ " loss.backward()\n",+ " return loss\n",+ " opt.step(closure)\n",+ " return float(torch.clamp(log_t.exp(), 0.1, 10.0).item())\n",+ "\n","def main():\n"," dist.init_process_group(\"nccl\")\n"," rank = dist.get_rank()\n",@@ -221,17 +240,17 @@" ddp_model = DDP(model, device_ids=[local_rank], find_unused_parameters=True)\n"," \n"," all_items = torch.load(\"/kaggle/working/train_items.pt\", weights_only=False)\n",- " # Split items evenly across ranks\n"," my_items = all_items[rank::world_size]\n"," \n",- " EPOCHS = 2\n",- " MICRO_BATCH = 10 # 10 sequences per forward pass per GPU\n",- " GRAD_ACCUM = 2 # Effective batch across 2 GPUs = 40 sequences\n",- " GROUP_SIZE = 4\n",- " LR_ENCODER = 1.5e-5\n",- " LR_HEAD = 1.5e-4\n",- " SIGMA_START = 0.5\n",- " SIGMA_END = 0.2\n",+ " # 4 Epochs with Cosine Annealing and soft-teacher guidance\n",+ " EPOCHS = 4\n",+ " MICRO_BATCH = 8 # 8 sequences per GPU: fits comfortably up to 1024 tokens!\n",+ " GRAD_ACCUM = 4 # Effective batch across 2 GPUs = 64 sequences (8 * 2 * 4)\n",+ " GROUP_SIZE = 4 # GRPO baseline samples\n",+ " LR_ENCODER = 2.0e-5 # Encoder adaptation rate\n",+ " LR_HEAD = 1.2e-4 # Head adaptation rate\n",+ " SIGMA_START = 0.5 # Exploration noise\n",+ " SIGMA_END = 0.15\n","\n"," enc_params = [p for n, p in ddp_model.named_parameters() if \"encoder.\" in n]\n"," head_params = [p for n, p in ddp_model.named_parameters() if \"encoder.\" not in n]\n",@@ -241,10 +260,12 @@" {\"params\": head_params, \"lr\": LR_HEAD}\n"," ], weight_decay=0.01)\n"," \n",+ " total_steps = (len(my_items) // (MICRO_BATCH * GRAD_ACCUM)) * EPOCHS\n",+ " scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max(1, total_steps), eta_min=1e-6)\n"," scaler = torch.amp.GradScaler(\"cuda\", enabled=True)\n"," \n"," if rank == 0:\n",- " print(f\"Starting 2xT4 DDP training: {len(all_items)} total items | {len(my_items)} per rank\")\n",+ " print(f\"Starting 2xT4 DDP training: {len(all_items)} total items | {len(my_items)} per rank | {EPOCHS} epochs\")\n"," t0 = time.time()\n"," \n"," for epoch in range(EPOCHS):\n",@@ -276,20 +297,25 @@" logits = logits.float()\n"," mask = batch[\"marker_mask\"].to(device)\n"," k = mask.sum(-1, keepdim=True).float()\n",+ " target = batch[\"target\"].to(device)\n"," \n",+ " # 1. Sample G noisy logit distributions with zero-mean projection\n"," eps = torch.randn((GROUP_SIZE,) + logits.shape, device=device) * sigma * mask\n"," eps = (eps - eps.sum(-1, keepdim=True) / k) * mask\n"," z = logits.detach().unsqueeze(0) + eps\n"," q = torch.softmax(z.masked_fill(~mask, -1e4), -1)\n"," \n",- " target = batch[\"target\"].to(device)\n",+ " # 2. Evaluate proper scoring reward (w_sph=0.75 for soft target matching)\n"," with torch.no_grad():\n",- " r = proper_reward(q, target.unsqueeze(0), batch[\"qtype\"].to(device), mask)\n",+ " r = proper_reward(q, target.unsqueeze(0), batch[\"qtype\"].to(device), mask, w_sph=0.75, w_rps=1.0)\n"," adv = r - r.mean(0, keepdim=True)\n"," adv = adv / (adv.std() + 1e-6)\n"," \n",+ " # 3. Policy gradient loss + auxiliary soft cross-entropy guidance (0.5 weight)\n"," logp = -(((z - logits.unsqueeze(0)) ** 2) * mask).sum(-1) / (2 * sigma ** 2)\n",- " loss = -(adv * logp).mean() / GRAD_ACCUM + 0.0 * act.sum()\n",+ " loss_rl = -(adv * logp).mean()\n",+ " loss_ce = -(target * torch.log_softmax(logits.masked_fill(~mask, -1e4), -1)).sum(-1).mean()\n",+ " loss = (loss_rl + 0.5 * loss_ce) / GRAD_ACCUM + 0.0 * act.sum()\n"," \n"," scaler.scale(loss).backward()\n"," accum_step += 1\n",@@ -299,26 +325,57 @@" torch.nn.utils.clip_grad_norm_(ddp_model.parameters(), 1.0)\n"," scaler.step(optimizer)\n"," scaler.update()\n",+ " scheduler.step()\n"," optimizer.zero_grad(set_to_none=True)\n"," \n"," epoch_loss += loss.item() * GRAD_ACCUM\n"," n_batches += 1\n"," \n"," if rank == 0 and (n_batches % 50) == 0:\n",- " print(f\" Epoch {epoch+1}/{EPOCHS} | Step {n_batches} | Loss: {loss.item()*GRAD_ACCUM:.4f} | Reward: {r.mean().item():.3f}\")\n",+ " cur_lr = scheduler.get_last_lr()[0]\n",+ " print(f\" Epoch {epoch+1}/{EPOCHS} | Step {n_batches} | Loss: {loss.item()*GRAD_ACCUM:.4f} | Reward: {r.mean().item():.3f} | LR: {cur_lr:.2e}\")\n","\n"," if rank == 0:\n"," print(f\"=== Epoch {epoch+1}/{EPOCHS} Completed in {time.time()-t0:.1f}s | Avg Loss: {epoch_loss/max(1, n_batches):.4f} ===\")\n","\n"," dist.barrier()\n",+ " \n",+ " # Post-training temperature calibration on rank 0\n"," if rank == 0:\n",+ " print(\"\\nFitting post-training calibration temperatures...\")\n",+ " model.eval()\n",+ " calib_items = all_items[:150]\n",+ " calib_batch = collate_train_batch(calib_items, tok.pad_token_id)\n",+ " with torch.no_grad():\n",+ " with torch.autocast(\"cuda\", dtype=torch.float16):\n",+ " logits_cal, _ = model(\n",+ " calib_batch[\"input_ids\"].to(device),\n",+ " calib_batch[\"attention_mask\"].to(device),\n",+ " calib_batch[\"marker_pos\"].to(device),\n",+ " calib_batch[\"marker_mask\"].to(device),\n",+ " calib_batch[\"qtype\"].to(device)\n",+ " )\n",+ " logits_cal = logits_cal.float().cpu().numpy()\n",+ " target_cal = calib_batch[\"target\"].numpy()\n",+ " qtypes_cal = calib_batch[\"qtype\"].numpy()\n",+ " \n",+ " fitted_temps = [1.0, 1.0, 1.0]\n",+ " for qt in range(3):\n",+ " sel = [(logits_cal[idx, :len(calib_items[idx][\"markers\"])], target_cal[idx, :len(calib_items[idx][\"markers\"])])\n",+ " for idx in range(len(calib_items)) if qtypes_cal[idx] == 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",+ " \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"," model.encoder.config.save_pretrained(os.path.join(output_dir, \"encoder\"))\n"," tok.save_pretrained(os.path.join(output_dir, \"tokenizer\"))\n",+ " \n"," cfg[\"fine_tuned\"] = True\n"," cfg[\"model_name\"] = \"laya-typed-decisions\"\n",+ " cfg[\"temperature\"] = fitted_temps\n"," with open(os.path.join(output_dir, \"rl_agent_config.json\"), \"w\") as f:\n"," json.dump(cfg, f, indent=2)\n"," print(f\"Model successfully saved to {output_dir}!\")\n",