Add mBERT-SV checkpoint-1072 (F1-macro 0.7777)

README.md

@@ -0,0 +1,99 @@
+---
+language:
+  - es
+license: apache-2.0
+base_model: google-bert/bert-base-multilingual-cased
+tags:
+  - text-classification
+  - hate-speech
+  - toxicity
+  - spanish
+  - el-salvador
+  - mbert
+pipeline_tag: text-classification
+library_name: transformers
+metrics:
+  - f1
+  - accuracy
+model-index:
+  - name: mbert-sv
+    results:
+      - task:
+          type: text-classification
+          name: Toxicidad en español salvadoreño
+        dataset:
+          name: Corpus tesina (test, n=460)
+          type: custom
+        metrics:
+          - type: f1
+            value: 0.7777
+            name: F1-macro
+          - type: accuracy
+            value: 0.7826
+            name: Accuracy
+---
+
+# mBERT-SV — Clasificación de toxicidad (El Salvador)
+
+Fine-tuning de [`bert-base-multilingual-cased`](https://huggingface.co/google-bert/bert-base-multilingual-cased) para clasificación **multiclase** de toxicidad en redes sociales en español salvadoreño.
+
+Checkpoint de auditoría: **`checkpoint-1072`** (época 4, semilla 42).
+
+## Etiquetas
+
+| id | etiqueta |
+|----|----------|
+| 0 | No Tóxico |
+| 1 | Lenguaje Ofensivo |
+| 2 | Discurso de Odio |
+| 3 | Amenazas/Violencia |
+
+## Contrato de inferencia
+
+| Campo | Valor |
+|-------|-------|
+| `max_length` | 128 |
+| Columna de texto | `texto_modelo` |
+| Normalizador | `normalize_for_model` v1.1 |
+
+Si el texto no está preprocesado, hay que aplicar el mismo normalizador del repositorio de la tesina.
+
+## Métricas oficiales (test, n = 460)
+
+| Métrica | Valor |
+|---------|-------|
+| F1-macro | **0.7777** |
+| Accuracy | 0.7826 |
+| Precision-macro | 0.7809 |
+| Recall-macro | 0.7757 |
+| QWK | 0.7362 |
+
+F1 por clase: No Tóxico 0.814 · Lenguaje Ofensivo 0.795 · Discurso de Odio 0.746 · Amenazas/Violencia 0.756.
+
+## Uso
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+repo_id = "caeher/mbert-sv"
+tokenizer = AutoTokenizer.from_pretrained(repo_id)
+model = AutoModelForSequenceClassification.from_pretrained(repo_id)
+
+text = "ejemplo de comentario"
+inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
+with torch.no_grad():
+    logits = model(**inputs).logits
+pred = int(logits.argmax(dim=-1))
+print(model.config.id2label[pred])
+```
+
+## Archivos
+
+- `model.safetensors`, `config.json`, `tokenizer.json`, `tokenizer_config.json`
+- `inference_contract.json` — contrato de preprocesamiento
+- `test_metrics.json` / `train_metrics.json` — métricas del checkpoint
+
+## Limitaciones
+
+Modelo de tesina, no un moderador autónomo. El corpus es de El Salvador; el desempeño puede caer en otros dialectos o en jerga adversarial. FPR sobre jerga no tóxica y cortes por plataforma están documentados en el informe del proyecto.

config.json

@@ -0,0 +1,46 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": null,
+  "classifier_dropout": null,
+  "directionality": "bidi",
+  "dtype": "float32",
+  "eos_token_id": null,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "No Tóxico",
+    "1": "Lenguaje Ofensivo",
+    "2": "Discurso de Odio",
+    "3": "Amenazas/Violencia"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "label2id": {
+    "No Tóxico": 0,
+    "Lenguaje Ofensivo": 1,
+    "Discurso de Odio": 2,
+    "Amenazas/Violencia": 3
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "pooler_fc_size": 768,
+  "pooler_num_attention_heads": 12,
+  "pooler_num_fc_layers": 3,
+  "pooler_size_per_head": 128,
+  "pooler_type": "first_token_transform",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.14.1",
+  "type_vocab_size": 2,
+  "use_cache": false,
+  "vocab_size": 119547
+}

inference_contract.json

@@ -0,0 +1,6 @@
+{
+  "max_length": 128,
+  "text_column": "texto_modelo",
+  "normalizer": "normalize_for_model",
+  "normalizer_version": "1.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2de91405b7c094368073b9b15e5a61287a987f7a4e308e1acd414dfbf36451f
+size 711449584

test_metrics.json

@@ -0,0 +1,38 @@
+{
+  "accuracy": 0.782608695652174,
+  "balanced_accuracy": 0.7757451391051162,
+  "f1_macro": 0.7777194765714894,
+  "f1_weighted": 0.7828842207289413,
+  "precision_macro": 0.7808732555624674,
+  "recall_macro": 0.7757451391051162,
+  "ordinal_mae": 0.3239130434782609,
+  "ordinal_rmse": 0.7561343335907864,
+  "large_error_pct": 8.91304347826087,
+  "qwk": 0.7361556982343499,
+  "kappa_linear": 0.7182345879992764,
+  "mcc": 0.7035845189708845,
+  "precision_No Tóxico": 0.8495575221238938,
+  "recall_No Tóxico": 0.7804878048780488,
+  "f1_No Tóxico": 0.8135593220338984,
+  "support_No Tóxico": 123,
+  "precision_Lenguaje Ofensivo": 0.7738095238095238,
+  "recall_Lenguaje Ofensivo": 0.8176100628930818,
+  "f1_Lenguaje Ofensivo": 0.7951070336391437,
+  "support_Lenguaje Ofensivo": 159,
+  "precision_Discurso de Odio": 0.7346938775510204,
+  "recall_Discurso de Odio": 0.7578947368421053,
+  "f1_Discurso de Odio": 0.7461139896373057,
+  "support_Discurso de Odio": 95,
+  "precision_Amenazas/Violencia": 0.7654320987654321,
+  "recall_Amenazas/Violencia": 0.7469879518072289,
+  "f1_Amenazas/Violencia": 0.7560975609756098,
+  "support_Amenazas/Violencia": 83,
+  "roc_auc_ovr_macro": 0.9221629865233141,
+  "log_loss": 0.8721387005957749,
+  "ece": 0.15694192051887512,
+  "brier_macro": 0.09395594522356987,
+  "classification_report": "                    precision    recall  f1-score   support\n\n         No Tóxico       0.85      0.78      0.81       123\n Lenguaje Ofensivo       0.77      0.82      0.80       159\n  Discurso de Odio       0.73      0.76      0.75        95\nAmenazas/Violencia       0.77      0.75      0.76        83\n\n          accuracy                           0.78       460\n         macro avg       0.78      0.78      0.78       460\n      weighted avg       0.78      0.78      0.78       460\n",
+  "f1_bootstrap_mean": 0.7775980516278476,
+  "f1_bootstrap_lo": 0.7357670523307825,
+  "f1_bootstrap_hi": 0.8143980633933287
+}

tokenizer_config.json

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

train_metrics.json

@@ -0,0 +1,9 @@
+{
+  "eval_loss": 0.8894229531288147,
+  "eval_f1_macro": 0.771461875533477,
+  "eval_accuracy": 0.7712418300653595,
+  "eval_runtime": 86.7344,
+  "eval_samples_per_second": 5.292,
+  "eval_steps_per_second": 0.334,
+  "epoch": 4.0
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be351f95ce0b99bab71f3fe9f100d435804bd29fc6ce4f857bce35a9a1603d7c
+size 5201