Upload folder using huggingface_hub

README.md

@@ -0,0 +1,116 @@
+---
+language: ar
+tags:
+- text-classification
+- arabic
+- levantine
+- marbert
+- hate-speech
+- incitement-detection
+- custom-code
+metrics:
+- accuracy
+- f1_macro
+- f1_incitement
+---
+
+# Levantine Arabic Incitement Detector (CLS + Mean + Max)
+
+This repository contains a custom fine-tuned **MARBERTv2** pooled-architecture model for 3-way classification of Levantine Arabic social text:
+
+- `normal`
+- `abusive`
+- `incitement`
+
+The training setup uses:
+- class-balanced cross-entropy
+- asymmetric error cost for incitement mistakes
+- a small ordinal penalty
+- an auxiliary lexicon head based on `incitement.csv`
+- pooled sentence representation: `CLS + mean pooling + max pooling`
+
+## Validation Summary
+
+These are the k-fold cross-validation averages used to select this configuration:
+
+| Metric | Value |
+| :-- | --: |
+| Accuracy | 82.40% |
+| F1 Macro | 0.8025 |
+| F1 Incitement | 0.7752 |
+
+## Labels
+
+| ID | Label |
+| :-- | :-- |
+| 0 | normal |
+| 1 | abusive |
+| 2 | incitement |
+
+## Confusion Matrix
+
+The image below is a **training-set sanity check** for the final model trained on all data. It is not an unbiased test result.
+
+![Confusion Matrix](./confusion_matrix.png)
+
+## How to Load
+
+This is a custom model wrapper, so load it with the provided `model.pt` weights plus the MARBERTv2 encoder and tokenizer.
+
+```python
+import json
+import os
+import torch
+import torch.nn as nn
+from transformers import AutoConfig, AutoModel, AutoTokenizer
+
+MODEL_DIR = "amitca71/marbertv2-levantine-incitement-detector-cls-mean-max"
+
+class MarbertMultiTask(nn.Module):
+    def __init__(self, model_name: str, pooling_strategy: str = "cls_mean_max"):
+        super().__init__()
+        config = AutoConfig.from_pretrained(model_name, num_labels=3)
+        self.encoder = AutoModel.from_pretrained(model_name, config=config)
+        self.pooling_strategy = pooling_strategy
+        hidden_size = self.encoder.config.hidden_size
+        self.dropout = nn.Dropout(getattr(self.encoder.config, "hidden_dropout_prob", 0.2))
+        self.classifier = nn.Linear(hidden_size * 3, 3)
+        self.lexicon_head = nn.Linear(hidden_size * 3, 1)
+
+    def _masked_mean(self, hidden_states, attention_mask):
+        mask = attention_mask.unsqueeze(-1).float()
+        masked = hidden_states * mask
+        denom = mask.sum(dim=1).clamp(min=1.0)
+        return masked.sum(dim=1) / denom
+
+    def _masked_max(self, hidden_states, attention_mask):
+        mask = attention_mask.unsqueeze(-1).bool()
+        masked = hidden_states.masked_fill(~mask, float("-inf"))
+        pooled = masked.max(dim=1).values
+        pooled[torch.isinf(pooled)] = 0.0
+        return pooled
+
+    def forward(self, input_ids, attention_mask, token_type_ids=None):
+        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
+        cls_state = outputs.last_hidden_state[:, 0]
+        mean_state = self._masked_mean(outputs.last_hidden_state, attention_mask)
+        max_state = self._masked_max(outputs.last_hidden_state, attention_mask)
+        pooled_state = self.dropout(torch.cat([cls_state, mean_state, max_state], dim=-1))
+        logits = self.classifier(pooled_state)
+        return logits
+
+model = MarbertMultiTask("UBC-NLP/MARBERTv2", pooling_strategy="cls_mean_max")
+state_dict = torch.load(os.path.join(MODEL_DIR, "model.pt"), map_location="cpu")
+model.load_state_dict(state_dict)
+model.eval()
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
+with open(os.path.join(MODEL_DIR, "label_map.json"), encoding="utf-8") as f:
+    label_map = json.load(f)
+```
+
+## Limitations
+
+- The top class is built from open-source proxies for incitement, not a gold incitement-only annotation project.
+- Performance is best on Levantine political/social text and may degrade on other Arabic varieties or platforms.
+- The confusion matrix in this card is from the full-data training run and should not be treated as held-out evaluation.

checkpoint-1748/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17490a58e11a2f0b010147faf35d59deb9213b71bd8e2b4232f19514292219ad
+size 651426560

checkpoint-1748/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ac3774da3dc92b6121a4beeae1e7a0f448dc53bc40a15e3da3a6cd06fab3b3c
+size 340377291

checkpoint-1748/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b85e51f1b32ba6c5e5ee05aef206f3d2a0346e1e2f6b7172203054c3aa73614
+size 14709

checkpoint-1748/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ad2e5787d1be4728ffadf823370726371ee5d4a156e2aba2c9dc2b59e032c2e
+size 1383

checkpoint-1748/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:929c9f4ef5e448da4bab7b0f36d9731ea203ded1e8d25831155becbbe0952e32
+size 1465

checkpoint-1748/trainer_state.json

@@ -0,0 +1,62 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 4.0,
+  "eval_steps": 500,
+  "global_step": 1748,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.0,
+      "grad_norm": 4.120875835418701,
+      "learning_rate": 4.170375079465988e-06,
+      "loss": 1.8289,
+      "step": 437
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 1.2741719484329224,
+      "learning_rate": 2.781309599491418e-06,
+      "loss": 1.287,
+      "step": 874
+    },
+    {
+      "epoch": 3.0,
+      "grad_norm": 2.3260374069213867,
+      "learning_rate": 1.3922441195168468e-06,
+      "loss": 1.2288,
+      "step": 1311
+    },
+    {
+      "epoch": 4.0,
+      "grad_norm": 26.597824096679688,
+      "learning_rate": 3.178639542275906e-09,
+      "loss": 1.2176,
+      "step": 1748
+    }
+  ],
+  "logging_steps": 500,
+  "max_steps": 1748,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-1748/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:274259ee33913acedabcfb1e8abe7fc4015bbb0f0dfc761cba9762cf9aa74bda
+size 5905

confusion_matrix.csv

@@ -0,0 +1,4 @@
+,pred_normal,pred_abusive,pred_incitement
+true_normal,6733,1069,75
+true_abusive,481,3948,468
+true_incitement,4,43,1132

label_map.json

@@ -0,0 +1,5 @@
+{
+  "0": "normal",
+  "1": "abusive",
+  "2": "incitement"
+}

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be519b32387d9988d14d71b5aef7510ed1a265eb3ddc7e1b85ea7df52b04dc2d
+size 651472903

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17490a58e11a2f0b010147faf35d59deb9213b71bd8e2b4232f19514292219ad
+size 651426560

special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

tokenizer_config.json

@@ -0,0 +1,58 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "never_split": null,
+  "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 sha256:274259ee33913acedabcfb1e8abe7fc4015bbb0f0dfc761cba9762cf9aa74bda
+size 5905

training_metadata.json

@@ -0,0 +1,47 @@
+{
+  "model_name": "UBC-NLP/MARBERTv2",
+  "max_length": 160,
+  "num_epochs": 4,
+  "learning_rate": 5e-06,
+  "train_batch_size": 16,
+  "gradient_accumulation_steps": 2,
+  "freeze_embeddings": true,
+  "freeze_bottom_n_layers": 6,
+  "pooling_strategy": "cls_mean_max",
+  "classifier_input_dim": 2304,
+  "class_weights": [
+    0.5904532074928284,
+    0.949765145778656,
+    3.944868564605713
+  ],
+  "cost_matrix": [
+    [
+      1.0,
+      1.0,
+      1.0
+    ],
+    [
+      1.0,
+      1.0,
+      1.0
+    ],
+    [
+      3.0,
+      2.0,
+      1.0
+    ]
+  ],
+  "cost_loss_weight": 0.75,
+  "ordinal_loss_weight": 0.2,
+  "lexicon_aux_loss_weight": 0.3,
+  "lexicon_false_negative_boost": 1.75,
+  "rows": 13953,
+  "label_counts": {
+    "normal": 7877,
+    "abusive": 4897,
+    "incitement": 1179
+  },
+  "kfold_accuracy": 0.824,
+  "kfold_f1_macro": 0.8025,
+  "kfold_f1_incitement": 0.7752
+}