Update README.md

README.md

@@ -14,18 +14,18 @@ pipeline_tag: token-classification
 
 This repository contains a **token-classification model** trained on the **DAMASHA-MAS** benchmark, introduced in:
 
-> **DAMASHA: Detecting AI in Mixed Adversarial Texts via Segmentation with Human-interpretable Attribution** :contentReference[oaicite:0]{index=0}  
+> **DAMASHA: Detecting AI in Mixed Adversarial Texts via Segmentation with Human-interpretable Attribution**  
 
 The model aims to **segment mixed human–AI text** at *token level* – i.e., decide for each token whether it was written by a *human* or an *LLM*, even under **syntactic adversarial attacks**.
 
 - **Base encoders:**  
-  - [`FacebookAI/roberta-base`](https://huggingface.co/FacebookAI/roberta-base) :contentReference[oaicite:1]{index=1}  
-  - [`answerdotai/ModernBERT-base`](https://huggingface.co/answerdotai/ModernBERT-base) :contentReference[oaicite:2]{index=2}  
-- **Architecture (high level):** RoBERTa + ModernBERT feature fusion → BiGRU + CRF with the **Info-Mask** gating mechanism from the paper. :contentReference[oaicite:3]{index=3}  
+  - [`FacebookAI/roberta-base`](https://huggingface.co/FacebookAI/roberta-base) 
+  - [`answerdotai/ModernBERT-base`](https://huggingface.co/answerdotai/ModernBERT-base) 
+- **Architecture (high level):** RoBERTa + ModernBERT feature fusion → BiGRU + CRF with the **Info-Mask** gating mechanism from the paper.  
 - **Task:** Token classification (binary authorship: human vs AI).  
 - **Language:** English  
 - **License (this model):** MIT  
-- **Training data license:** CC-BY-4.0 via the DAMASHA dataset. :contentReference[oaicite:4]{index=4}  
+- **Training data license:** CC-BY-4.0 via the DAMASHA dataset.  
 
 If you use this model, **please also cite the DAMASHA paper and dataset** (see Citation section).
 
@@ -34,16 +34,16 @@ If you use this model, **please also cite the DAMASHA paper and dataset** (see C
 ## 1. Model Highlights
 
 - **Fine-grained mixed-authorship detection**  
-  Predicts authorship **per token**, allowing reconstruction of human vs AI **spans** in long documents. :contentReference[oaicite:5]{index=5}  
+  Predicts authorship **per token**, allowing reconstruction of human vs AI **spans** in long documents.  
 
 - **Adversarially robust**  
-  Trained and evaluated on **syntactically attacked texts** (misspelling, Unicode substitutions, invisible characters, punctuation swaps, case perturbations, and “all-mixed” attacks). :contentReference[oaicite:6]{index=6}  
+  Trained and evaluated on **syntactically attacked texts** (misspelling, Unicode substitutions, invisible characters, punctuation swaps, case perturbations, and “all-mixed” attacks).  
 
 - **Human-interpretable Info-Mask**  
-  The architecture incorporates **stylometric features** (perplexity, POS density, punctuation density, lexical diversity, readability) via an **Info-Mask** module that gates token representations in an interpretable way. :contentReference[oaicite:7]{index=7}  
+  The architecture incorporates **stylometric features** (perplexity, POS density, punctuation density, lexical diversity, readability) via an **Info-Mask** module that gates token representations in an interpretable way. 
 
 - **Strong reported performance (from the paper)**  
-  On DAMASHA-MAS, the **RMC\*** model (RoBERTa + ModernBERT + CRF + Info-Mask) achieves: :contentReference[oaicite:8]{index=8}  
+  On DAMASHA-MAS, the **RMC\*** model (RoBERTa + ModernBERT + CRF + Info-Mask) achieves:  
   - **Token-level**: Accuracy / Precision / Recall / F1 ≈ **0.98**  
   - **Span-level (strict)**: SBDA ≈ **0.45**, SegPre ≈ **0.41**  
   - **Span-level (relaxed IoU ≥ 0.5)**: ≈ **0.82**  
@@ -68,7 +68,7 @@ If you use this model, **please also cite the DAMASHA paper and dataset** (see C
 
 - Automated “cheating detector” / plagiarism court.  
 - High-stakes decisions affecting people’s livelihood, grades, or reputation **without human review**.  
-- Non-English or heavily code-mixed text (training data is English-centric). :contentReference[oaicite:9]{index=9}  
+- Non-English or heavily code-mixed text (training data is English-centric). 
 
 Use this model as a **signal**, not a judge.
 
@@ -78,11 +78,11 @@ Use this model as a **signal**, not a judge.
 
 The model is trained on the **MAS** benchmark released with the DAMASHA paper and hosted as the Hugging Face dataset:
 
-- **Dataset:** [`saiteja33/DAMASHA`](https://huggingface.co/datasets/saiteja33/DAMASHA) :contentReference[oaicite:10]{index=10}  
+- **Dataset:** [`saiteja33/DAMASHA`](https://huggingface.co/datasets/saiteja33/DAMASHA) 
 
 ### 3.1 What’s in MAS?
 
-MAS consists of **mixed human–AI texts with explicit span tags**: :contentReference[oaicite:11]{index=11}  
+MAS consists of **mixed human–AI texts with explicit span tags**:  
 
 - Human text comes from several corpora for **domain diversity**, including:  
   - Reddit (M4-Reddit)  
@@ -94,11 +94,11 @@ MAS consists of **mixed human–AI texts with explicit span tags**: :contentRefe
 
 - AI text is generated by multiple modern LLMs:  
   - **DeepSeek-V3-671B** (open-source)  
-  - **GPT-4o, GPT-4.1, GPT-4.1-mini** (closed-source) :contentReference[oaicite:12]{index=12}  
+  - **GPT-4o, GPT-4.1, GPT-4.1-mini** (closed-source)  
 
 ### 3.2 Span tagging
 
-Authorship is marked using **explicit tags** around AI spans: :contentReference[oaicite:13]{index=13}  
+Authorship is marked using **explicit tags** around AI spans:  
 
 - `<AI_Start>` … `</AI_End>` denote AI-generated segments within otherwise human text.  
 - The dataset stores text in a `hybrid_text` column, plus metadata such as `has_pair`, and adversarial variants include `attack_name`, `tag_count`, and `attacked_text`.  
@@ -108,7 +108,7 @@ Authorship is marked using **explicit tags** around AI spans: :contentReference[
 
 ### 3.3 Adversarial attacks
 
-MAS includes multiple **syntactic attacks** applied to the mixed text: :contentReference[oaicite:14]{index=14}  
+MAS includes multiple **syntactic attacks** applied to the mixed text:  
 
 - Misspelling  
 - Unicode character substitution  
@@ -125,7 +125,7 @@ These perturbations make tokenization brittle and test robustness of detectors i
 
 ### 4.1 Architecture (conceptual)
 
-The model follows the **Info-Mask RMC\*** architecture described in the DAMASHA paper: :contentReference[oaicite:15]{index=15}  
+The model follows the **Info-Mask RMC\*** architecture described in the DAMASHA paper:  
 
 1. **Dual encoders**  
    - RoBERTa-base and ModernBERT-base encode the same input sequence.  
@@ -139,7 +139,7 @@ The model follows the **Info-Mask RMC\*** architecture described in the DAMASHA
 
 ### 4.2 Training setup (from the paper)
 
-Key hyperparameters used for the Info-Mask models on MAS: :contentReference[oaicite:18]{index=18}  
+Key hyperparameters used for the Info-Mask models on MAS:  
 
 - **Number of labels:** 2  
 - **Max sequence length:** 512  
@@ -152,10 +152,10 @@ Key hyperparameters used for the Info-Mask models on MAS: :contentReference[oaic
 - **Warmup ratio:** 0.1  
 - **Early stopping patience:** 2  
 
-**Hardware & compute** (as reported): :contentReference[oaicite:19]{index=19}  
+**Hardware & compute** (as reported): 
 
 - AWS EC2 g6e.xlarge, NVIDIA L40S (48GB) GPU, Ubuntu 24.04  
-- ≈ 400 GPU hours (~USD $720) for experiments.
+- ≈ 400 GPU hours for experiments.
 
 > The exact training script used for this checkpoint is available in the project GitHub:  
 > <https://github.com/saitejalekkala33/DAMASHA>