Upload README.md with huggingface_hub

README.md

@@ -95,64 +95,6 @@ model.eval()
 
 Inputs are tokenised at a maximum sequence length of 100.
 
-## Reported results
-
-From the paper, on the test split as it stood at publication:
-
-| Model | Precision (%) | Recall (%) | F1 (%) |
-|---|---:|---:|---:|
-| **XLM-RoBERTa** | **95.22** | **95.38** | **95.25** |
-| BanglaBERT | 94.73 | 94.67 | 94.63 |
-| mBERT | 93.77 | 93.66 | 93.59 |
-| Logistic Regression | 86.31 | 85.02 | 84.04 |
-| K-Nearest Neighbors | 83.80 | 82.83 | 82.01 |
-| Multinomial Naive Bayes | 74.17 | 66.07 | 60.76 |
-| BiLSTM + CNN | 62.42 | 64.65 | 60.37 |
-| LSTM | 20.24 | 44.99 | 27.92 |
-
-Best results for XLM-RoBERTa came from batch size 16 and learning rate 1e-5;
-mBERT and BanglaBERT used batch size 16 and learning rate 3e-5.
-
-## Limitations
-
-Please read these before relying on any of the non-transformer models.
-
-**The published metrics were measured on an earlier version of the dataset.**
-That version contained 5,629 posts and included only 13 `blood` examples — 7 in
-train, 1 in validation, 5 in test. The `blood` column of the published confusion
-matrix therefore rests on five test samples. The dataset has since been corrected
-to the full 220 `blood` posts described above. These checkpoints were trained
-before that correction, so their real-world `blood` performance is far weaker
-than the aggregate scores suggest, and the table above does not describe
-performance on the current data.
-
-**The DNN and scikit-learn models are not self-contained.** They consume token
-ids from a Keras `Tokenizer` and TF-IDF features fitted at import time on the
-training CSV, rather than a saved vocabulary. Their embedding indices and feature
-columns are only meaningful against that exact fitted vocabulary. Reconstructing
-it from corrected data shifts the indices and the predictions become unreliable
-while still looking confident. Treat these six as historical baselines.
-
-**The scikit-learn models were pickled under scikit-learn 1.0.2.** Loading them
-on a later version raises `InconsistentVersionWarning`, which scikit-learn notes
-may lead to invalid results.
-
-**The Multinomial Naive Bayes model predicts string labels** (`'crime'`), while
-the logistic regression and KNN models predict integer ids. Callers must handle
-both.
-
-**Loading pickles executes arbitrary code.** The `.pkl` files and `torch.load`
-both unpickle. Only load them if you trust this repository.
-
-**Class imbalance and annotation noise.** `crime` dominates at 42.7%. The source
-spreadsheet also contains a small number of duplicate posts, four of which carry
-conflicting labels (for example the same fire report labelled both `fire` and
-`natural_disaster`) — a reminder that some categories genuinely overlap.
-
-**Scope.** Trained on Bangla posts about events in Bangladesh and neighbouring
-India. Behaviour on other varieties of Bangla, or on non-emergency text, is
-untested. This is a research artifact and is not fit to be a sole trigger for
-any real emergency response.
 
 ## Citation
 
@@ -168,9 +110,3 @@ any real emergency response.
   year      = {2023}
 }
 ```
-
-## Acknowledgements
-
-Department of Computer Science and Engineering, Khulna University of Engineering
-& Technology (KUET). The training code is derived from
-[xashru/bangla-text-classification](https://github.com/xashru/bangla-text-classification) (MIT).