feat: Turkish Financial BERT Router (3-class MoE)

README.md

@@ -0,0 +1,71 @@
+---
+language:
+- tr
+license: apache-2.0
+tags:
+- bert
+- turkish
+- finance
+- text-classification
+- router
+- query-classification
+library_name: transformers
+base_model: Toxotes/fin-bert-tr
+---
+
+# fin-bert-tr-router
+
+**Türkçe Finansal BERT Router** — Kullanıcı sorgularını 3 sınıfa ayıran
+sınıflandırma modeli. MOSAIC RAG sisteminde MoE (Mixture of Experts) yönlendiricisi
+olarak kullanılır.
+
+## Sınıflar
+
+| ID | Sınıf | Açıklama | Yönlendirme |
+|----|-------|----------|-------------|
+| 0  | CHITCHAT | Selamlama, sohbet, kimlik soruları | RAG olmadan Qwen2.5-1.5B |
+| 1  | SIMPLE | Tek adımlı olgusal finans sorguları | Qwen2.5-1.5B + RAG |
+| 2  | COMPLEX | Çok adımlı analiz, mevzuat yorumu | Turkish-Gemma-9B + RAG |
+
+## Performans
+
+- **Doğrulama Accuracy:** 0.8940
+- **Doğrulama Macro F1:** 0.9238
+
+## Temel Model
+
+[Toxotes/fin-bert-tr](https://huggingface.co/Toxotes/fin-bert-tr) üzerinde ince-ayar yapılmıştır.
+
+## Kullanım
+
+```python
+from transformers import BertForSequenceClassification, PreTrainedTokenizerFast
+import torch
+
+tokenizer = PreTrainedTokenizerFast.from_pretrained("Toxotes/fin-bert-tr-router")
+model = BertForSequenceClassification.from_pretrained("Toxotes/fin-bert-tr-router")
+model.eval()
+
+queries = [
+    "Merhaba!",                                           # → CHITCHAT (0)
+    "TCMB faiz oranı nedir?",                             # → SIMPLE   (1)
+    "Reel sektör sermaye yeterliliği analizi nasıl yapılır?",  # → COMPLEX  (2)
+]
+LABELS = ["chitchat", "simple", "complex"]
+
+for q in queries:
+    inputs = tokenizer(q, return_tensors="pt", truncation=True, max_length=128)
+    with torch.no_grad():
+        logits = model(**inputs).logits
+    label = LABELS[logits.argmax().item()]
+    print(f"{q[:40]:40} → {label}")
+```
+
+## Eğitim Tarihi
+
+2026-03-11
+
+## Proje
+
+MOSAIC — Federated Financial RAG System
+[GitHub](https://github.com/tahatoy/MOSAIC)

checkpoint-180/config.json

@@ -0,0 +1,39 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "chitchat",
+    "1": "simple",
+    "2": "complex"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "chitchat": 0,
+    "complex": 2,
+    "simple": 1
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "problem_type": "single_label_classification",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.3.0",
+  "type_vocab_size": 2,
+  "use_cache": false,
+  "vocab_size": 32000
+}

checkpoint-180/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a67357f153d26a876ac69d386fd49fbf1abe99cf1f69b43efc7eb9d74988fedd
+size 442502116

checkpoint-180/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d472268a96b6545b8ac8ea1c1df0c6d30973d204ecc02896b3db19468cf42b03
+size 885125771

checkpoint-180/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a727d9b55b2ea15801f73de2d67be739cb2afc7490ba9ff2a93d2fd0c1b4a0c
+size 14645

checkpoint-180/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:343bc140b7878c594877e97f280bb1571cb798eaae931c6ef019fd867e4b99b2
+size 1383

checkpoint-180/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d4a004c920138b162037c9e97d79bd1de73e6cd668c54aed9bc3c469af97dd0
+size 1465

checkpoint-180/trainer_state.json

@@ -0,0 +1,116 @@
+{
+  "best_global_step": 180,
+  "best_metric": 0.9214996490839781,
+  "best_model_checkpoint": "C:\\Users\\memol\\Desktop\\Kisisel_Projeler\\MOSAIC\\federafinrag\\models\\router_bert\\checkpoint-180",
+  "epoch": 4.0,
+  "eval_steps": 500,
+  "global_step": 180,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.0,
+      "eval_accuracy": 0.7927215189873418,
+      "eval_chitchat_f1": 0.7897271268057785,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.4422382414340973,
+      "eval_macro_f1": 0.8503704873202018,
+      "eval_runtime": 0.5699,
+      "eval_samples_per_second": 1108.897,
+      "eval_simple_f1": 0.761384335154827,
+      "eval_steps_per_second": 8.773,
+      "step": 45
+    },
+    {
+      "epoch": 1.1111111111111112,
+      "grad_norm": 2.4390697479248047,
+      "learning_rate": 1.909342403345925e-05,
+      "loss": 0.6157618713378906,
+      "step": 50
+    },
+    {
+      "epoch": 2.0,
+      "eval_accuracy": 0.870253164556962,
+      "eval_chitchat_f1": 0.8624161073825504,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.3124813735485077,
+      "eval_macro_f1": 0.9066849987571465,
+      "eval_runtime": 0.5695,
+      "eval_samples_per_second": 1109.712,
+      "eval_simple_f1": 0.8576388888888888,
+      "eval_steps_per_second": 8.779,
+      "step": 90
+    },
+    {
+      "epoch": 2.2222222222222223,
+      "grad_norm": 2.671827554702759,
+      "learning_rate": 1.3666992531768482e-05,
+      "loss": 0.34577438354492185,
+      "step": 100
+    },
+    {
+      "epoch": 3.0,
+      "eval_accuracy": 0.8860759493670886,
+      "eval_chitchat_f1": 0.8783783783783784,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.2773128151893616,
+      "eval_macro_f1": 0.9180801491146319,
+      "eval_runtime": 0.5695,
+      "eval_samples_per_second": 1109.75,
+      "eval_simple_f1": 0.8758620689655172,
+      "eval_steps_per_second": 8.78,
+      "step": 135
+    },
+    {
+      "epoch": 3.3333333333333335,
+      "grad_norm": 3.06937313079834,
+      "learning_rate": 6.1908705719647735e-06,
+      "loss": 0.2207444953918457,
+      "step": 150
+    },
+    {
+      "epoch": 4.0,
+      "eval_accuracy": 0.8908227848101266,
+      "eval_chitchat_f1": 0.8828522920203735,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.27712756395339966,
+      "eval_macro_f1": 0.9214996490839781,
+      "eval_runtime": 0.5668,
+      "eval_samples_per_second": 1115.105,
+      "eval_simple_f1": 0.8816466552315609,
+      "eval_steps_per_second": 8.822,
+      "step": 180
+    }
+  ],
+  "logging_steps": 50,
+  "max_steps": 225,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "EarlyStoppingCallback": {
+      "args": {
+        "early_stopping_patience": 3,
+        "early_stopping_threshold": 0.0
+      },
+      "attributes": {
+        "early_stopping_patience_counter": 0
+      }
+    },
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1495273552948224.0,
+  "train_batch_size": 128,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-180/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60ca23d105e7290f4c5f2500d68d89996b07cd664a889d31bd6c74f4043d1f02
+size 5265

checkpoint-225/config.json

@@ -0,0 +1,39 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "chitchat",
+    "1": "simple",
+    "2": "complex"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "chitchat": 0,
+    "complex": 2,
+    "simple": 1
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "problem_type": "single_label_classification",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.3.0",
+  "type_vocab_size": 2,
+  "use_cache": false,
+  "vocab_size": 32000
+}

checkpoint-225/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d15e24d35d8168667200ec1fdbf85419fb9ba4ac4ee384e91cdcc1c06222e86
+size 442502116

checkpoint-225/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e13e7cbbad2737a9fbe417ce50e8be8108235831cb7653c10f5592254f8542d
+size 885125771

checkpoint-225/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8b6282e3b9f4ef473373e2fff949dbe62945fd35d70c4738f5fc1af2f5039e2
+size 14645

checkpoint-225/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f53e6e635233bdd30b27e3b4e224f937719c4918e21e112f405f94670790aa4d
+size 1383

checkpoint-225/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b396cd839a80b567f2098e9c2387f4ec8ce92c8eecf966902cc42b39db47c413
+size 1465

checkpoint-225/trainer_state.json

@@ -0,0 +1,136 @@
+{
+  "best_global_step": 225,
+  "best_metric": 0.9237714697319616,
+  "best_model_checkpoint": "C:\\Users\\memol\\Desktop\\Kisisel_Projeler\\MOSAIC\\federafinrag\\models\\router_bert\\checkpoint-225",
+  "epoch": 5.0,
+  "eval_steps": 500,
+  "global_step": 225,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.0,
+      "eval_accuracy": 0.7927215189873418,
+      "eval_chitchat_f1": 0.7897271268057785,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.4422382414340973,
+      "eval_macro_f1": 0.8503704873202018,
+      "eval_runtime": 0.5699,
+      "eval_samples_per_second": 1108.897,
+      "eval_simple_f1": 0.761384335154827,
+      "eval_steps_per_second": 8.773,
+      "step": 45
+    },
+    {
+      "epoch": 1.1111111111111112,
+      "grad_norm": 2.4390697479248047,
+      "learning_rate": 1.909342403345925e-05,
+      "loss": 0.6157618713378906,
+      "step": 50
+    },
+    {
+      "epoch": 2.0,
+      "eval_accuracy": 0.870253164556962,
+      "eval_chitchat_f1": 0.8624161073825504,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.3124813735485077,
+      "eval_macro_f1": 0.9066849987571465,
+      "eval_runtime": 0.5695,
+      "eval_samples_per_second": 1109.712,
+      "eval_simple_f1": 0.8576388888888888,
+      "eval_steps_per_second": 8.779,
+      "step": 90
+    },
+    {
+      "epoch": 2.2222222222222223,
+      "grad_norm": 2.671827554702759,
+      "learning_rate": 1.3666992531768482e-05,
+      "loss": 0.34577438354492185,
+      "step": 100
+    },
+    {
+      "epoch": 3.0,
+      "eval_accuracy": 0.8860759493670886,
+      "eval_chitchat_f1": 0.8783783783783784,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.2773128151893616,
+      "eval_macro_f1": 0.9180801491146319,
+      "eval_runtime": 0.5695,
+      "eval_samples_per_second": 1109.75,
+      "eval_simple_f1": 0.8758620689655172,
+      "eval_steps_per_second": 8.78,
+      "step": 135
+    },
+    {
+      "epoch": 3.3333333333333335,
+      "grad_norm": 3.06937313079834,
+      "learning_rate": 6.1908705719647735e-06,
+      "loss": 0.2207444953918457,
+      "step": 150
+    },
+    {
+      "epoch": 4.0,
+      "eval_accuracy": 0.8908227848101266,
+      "eval_chitchat_f1": 0.8828522920203735,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.27712756395339966,
+      "eval_macro_f1": 0.9214996490839781,
+      "eval_runtime": 0.5668,
+      "eval_samples_per_second": 1115.105,
+      "eval_simple_f1": 0.8816466552315609,
+      "eval_steps_per_second": 8.822,
+      "step": 180
+    },
+    {
+      "epoch": 4.444444444444445,
+      "grad_norm": 2.217465877532959,
+      "learning_rate": 8.438874567923261e-07,
+      "loss": 0.16295211791992187,
+      "step": 200
+    },
+    {
+      "epoch": 5.0,
+      "eval_accuracy": 0.8939873417721519,
+      "eval_chitchat_f1": 0.8866328257191202,
+      "eval_complex_f1": 1.0,
+      "eval_loss": 0.2775741517543793,
+      "eval_macro_f1": 0.9237714697319616,
+      "eval_runtime": 0.5907,
+      "eval_samples_per_second": 1069.839,
+      "eval_simple_f1": 0.8846815834767642,
+      "eval_steps_per_second": 8.464,
+      "step": 225
+    }
+  ],
+  "logging_steps": 50,
+  "max_steps": 225,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "EarlyStoppingCallback": {
+      "args": {
+        "early_stopping_patience": 3,
+        "early_stopping_threshold": 0.0
+      },
+      "attributes": {
+        "early_stopping_patience_counter": 0
+      }
+    },
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1869091941185280.0,
+  "train_batch_size": 128,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-225/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60ca23d105e7290f4c5f2500d68d89996b07cd664a889d31bd6c74f4043d1f02
+size 5265

classification_report.txt

@@ -0,0 +1,17 @@
+=== Classification Report ===
+              precision    recall  f1-score   support
+
+    chitchat     0.8822    0.8912    0.8866       294
+      simple     0.8893    0.8801    0.8847       292
+     complex     1.0000    1.0000    1.0000        46
+
+    accuracy                         0.8940       632
+   macro avg     0.9238    0.9238    0.9238       632
+weighted avg     0.8940    0.8940    0.8940       632
+
+
+=== Confusion Matrix ===
+  Rows: truth, Cols: pred, Labels: ['chitchat', 'simple', 'complex']
+[[262  32   0]
+ [ 35 257   0]
+ [  0   0  46]]

config.json

@@ -0,0 +1,39 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "chitchat",
+    "1": "simple",
+    "2": "complex"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "chitchat": 0,
+    "complex": 2,
+    "simple": 1
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "problem_type": "single_label_classification",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.3.0",
+  "type_vocab_size": 2,
+  "use_cache": false,
+  "vocab_size": 32000
+}

meta.json

@@ -0,0 +1,14 @@
+{
+  "base_model": "dbmdz/bert-base-turkish-cased",
+  "architecture": "AutoModelForSequenceClassification → 3-class router",
+  "num_labels": 3,
+  "label_names": [
+    "chitchat",
+    "simple",
+    "complex"
+  ],
+  "val_accuracy": 0.8939873417721519,
+  "val_macro_f1": 0.9237714697319616,
+  "n_train": 5683,
+  "n_val": 632
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d15e24d35d8168667200ec1fdbf85419fb9ba4ac4ee384e91cdcc1c06222e86
+size 442502116

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "backend": "tokenizers",
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "is_local": false,
+  "mask_token": "[MASK]",
+  "max_len": 512,
+  "model_max_length": 512,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]"
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60ca23d105e7290f4c5f2500d68d89996b07cd664a889d31bd6c74f4043d1f02
+size 5265