Upload folder using huggingface_hub

README.md

@@ -1,71 +1,157 @@
-# DeBERTa Multi-Label Financial Event Classifier (SENTiVENT)
-
-This model is a fine-tuned **DeBERTa** classifier that predicts one or more **financial event types** from a news headline or short piece of text.
-
-Given a headline, the model outputs probabilities for multiple event categories such as mergers, earnings reports, legal actions, investments, and more. It is designed for **multi-label classification**, meaning a single headline can belong to multiple event types at once.
-
+---
+language:
+- en
+license: mit
+tags:
+- text-classification
+- multi-label-classification
+- financial-nlp
+- finance
+- event-detection
+datasets:
+- sentivent
+metrics:
+- f1
+- precision
+- recall
+pipeline_tag: text-classification
 ---
 
-## Event Labels
-
-The model predicts the following 18 event categories:
-
-- CSR/Brand  
-- Deal  
-- Dividend  
-- Employment  
-- Expense  
-- Facility  
-- FinancialReport  
-- Financing  
-- Investment  
-- Legal  
-- Macroeconomics  
-- Merger/Acquisition  
-- Product/Service  
-- Profit/Loss  
-- Rating  
-- Revenue  
-- SalesVolume  
-- SecurityValue  
+# FinDeBERTa: Multi-Label Financial Event Classifier
 
----
+FinDeBERTa is a fine-tuned DeBERTa-v3-Large model for multi-label financial event classification. It predicts one or more event types from financial news headlines with state-of-the-art performance.
 
-## Intended Use
+## Model Details
 
-This model is useful for:
+- **Base Model**: microsoft/deberta-v3-large
+- **Task**: Multi-label text classification
+- **Training**: Fine-tuned with Focal Loss and per-class threshold optimization
+- **Performance**: Macro F1: 0.686 | Precision: 0.731 | Recall: 0.685
 
-- Financial news analysis
-- Market event extraction
-- Trading signal pipelines
-- Knowledge graph population
-- Research in finance-focused NLP
+## Event Labels
 
-It works best on **short financial news headlines or sentences**.
+The model classifies text into 18 financial event types:
 
----
+```python
+["CSR/Brand", "Deal", "Dividend", "Employment", "Expense", "Facility",
+ "FinancialReport", "Financing", "Investment", "Legal", "Macroeconomics",
+ "Merger/Acquisition", "Product/Service", "Profit/Loss", "Rating", "Revenue",
+ "SalesVolume", "SecurityValue"]
+```
 
-## Model Details
+## Usage
+
+### Basic Usage (with default threshold)
 
-- Base model: `deberta-v3-large`
-- Task: Multi-label text classification
-- Activation: Sigmoid (per-label probabilities)
-- Loss: Binary Cross Entropy
-- Problem type: `multi_label_classification`
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+import numpy as np
 
-The model configuration includes `id2label`, `label2id`, and the correct problem type so it works cleanly with Hugging Face pipelines.
+model_name = "ritessshhh/FinDeBERTa"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
 
----
-license: mit
-language:
-- en
-metrics:
-- f1
-- precision
-- recall
-base_model:
-- microsoft/deberta-v3-large
-pipeline_tag: text-classification
-tags:
-- finance
----
+text = "Tesla to acquire a battery startup in a 400 million dollar deal."
+inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    probs = torch.sigmoid(outputs.logits)[0].cpu().numpy()
+
+# Using default threshold of 0.5
+threshold = 0.5
+predictions = [
+    {"label": model.config.id2label[i], "score": float(prob)}
+    for i, prob in enumerate(probs) if prob >= threshold
+]
+
+print(predictions)
+```
+
+### Optimized Usage (with per-class thresholds)
+
+For best performance, use the per-class optimized thresholds:
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+import numpy as np
+from huggingface_hub import hf_hub_download
+
+model_name = "ritessshhh/FinDeBERTa"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+
+# Download per-class thresholds
+thresholds_path = hf_hub_download(repo_id=model_name, filename="thresholds.npy")
+thresholds = np.load(thresholds_path)
+
+text = "Tesla to acquire a battery startup in a 400 million dollar deal."
+inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    probs = torch.sigmoid(outputs.logits)[0].cpu().numpy()
+
+# Apply per-class thresholds
+predictions = [
+    {"label": model.config.id2label[i], "score": float(prob)}
+    for i, prob in enumerate(probs) if prob >= thresholds[i]
+]
+
+# Sort by score
+predictions = sorted(predictions, key=lambda x: x["score"], reverse=True)
+print(predictions)
+# Output: [{"label": "Merger/Acquisition", "score": 0.98}, ...]
+```
+
+## Training Details
+
+- **Loss Function**: Focal Loss (gamma=2.0) with dampened class weights
+- **Optimizer**: AdamW with cosine learning rate scheduling
+- **Batch Size**: 8 (with gradient accumulation steps=2)
+- **Epochs**: 10
+- **Learning Rate**: 2e-5
+- **Weight Decay**: 0.02
+
+## Performance
+
+| Metric | Score |
+|--------|-------|
+| Macro F1 | 0.686 |
+| Micro F1 | 0.710 |
+| Precision (Macro) | 0.731 |
+| Recall (Macro) | 0.685 |
+| Exact Match Ratio | 0.569 |
+
+### Per-Class Performance
+
+| Label | F1 Score |
+|-------|----------|
+| Merger/Acquisition | 0.896 |
+| Dividend | 0.923 |
+| Profit/Loss | 0.863 |
+| Employment | 0.833 |
+| Rating | 0.821 |
+| SalesVolume | 0.820 |
+
+## Limitations
+
+- Optimized for financial news headlines (short text)
+- May not generalize well to other domains
+- Performance varies by event type (rare events like "Facility" have lower F1)
+
+## Citation
+
+If you use this model, please cite:
+
+```bibtex
+@misc{findeberta2024,
+  author = {ritessshhh},
+  title = {FinDeBERTa: Multi-Label Financial Event Classifier},
+  year = {2024},
+  publisher = {HuggingFace},
+  howpublished = {\url{https://huggingface.co/ritessshhh/FinDeBERTa}}
+}
+```