Update README.md

README.md

@@ -40,7 +40,6 @@ The model was trained on a curated and annotated corpus of official municipal me
 - **Task:** Abstractive summarization (`text → summary`)  
 - **Framework:** 🤗 Transformers (PyTorch)  
 - **Max Input Length:** 512 tokens  
-- **Max Summary Length:** 128 tokens  
 - **Training Objective:** Conditional generation (cross-entropy loss)  
 - **Dataset:** Portuguese municipal meeting minutes annotated with summaries  
 
@@ -58,7 +57,7 @@ The model receives a discussion subject of a municipal meeting and outputs a sho
 ```python
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 
-model_name = "anonymous12321/CitilinkSumm-PT"
+model_name = "anonymous12321/Pegasus-Summarization-Council-PT"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
 
@@ -85,26 +84,36 @@ print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))
 
 | Metric | Score | Description |
 |:-------|:------:|:------------|
-| **ROUGE-1** | ... | Unigram overlap between generated and reference summaries |
-| **ROUGE-2** | ... | Bigram overlap |
-| **ROUGE-L** | ... | Longest common subsequence overlap |
-| **BERTScore (F1)** | ... | Semantic similarity between summary and reference |
+| **ROUGE-1** | 0.367 | Unigram overlap between generated and reference summaries |
+| **ROUGE-2** | 0.179 | Bigram overlap |
+| **ROUGE-L** | 0.309 | Longest common subsequence overlap |
+| **BERTScore (F1)** | 0.746 | Semantic similarity between summary and reference |
 
 ---
 
 ## ⚙️ Training Details
 
 - **Pretrained Model:** `google/pegasus-xsum`  
+- **Tokenizer:** `AutoTokenizer` (matching checkpoint)  
 - **Optimizer:** AdamW (default in Hugging Face Trainer)  
 - **Learning Rate:** 2e-5  
 - **Batch Size:** 4  
 - **Epochs:** 3  
-- **Scheduler:** Linear warmup  
-- **Loss Function:** Cross-entropy  
+- **Scheduler:** Linear warmup (default)  
+- **Loss Function:** Cross-entropy (seq2seq objective)  
 - **Evaluation Metrics:** ROUGE (computed on validation set every 100 steps)  
-- **Evaluation Strategy:** Step-based evaluation (`eval_steps=100`)  
+- **Evaluation Strategy:** Step-based evaluation (`eval_strategy="steps"`, `eval_steps=100`)  
 - **Weight Decay:** 0.01  
+- **Logging Steps:** 10  
 - **Mixed Precision (fp16):** Enabled when CUDA is available  
+- **Save Strategy:** Keep only latest checkpoint (`save_total_limit=1`)  
+- **Chunking:** Token-based with `max_length=512` and `stride=256`  
+- **Target Max Length:** 128  
+- **Validation Split:** 10% of data  
+- **Data Collator:** `DataCollatorForSeq2Seq` (dynamic padding)  
+- **Output Directory:** `./results_hierarchical_pegasus_segments`  
+- **Saved Model Path:** `./trained_pegasus_segments`  
+
 
 ---
 
@@ -131,15 +140,6 @@ The model was trained on a specialized dataset of **Portuguese municipal meeting
 
 ---
 
-## ⚖️ Ethical Considerations
-
-The model is intended for **research and administrative document processing**.
-
-- Outputs should **not** be used for legal decision-making without human verification.  
-- Potential bias may exist due to limited geographic and institutional diversity in training data.  
-
----
-
 ## 📄 License
 
 This model is released under the