Upload README.md with huggingface_hub

README.md

@@ -7,7 +7,6 @@ library_name: transformers
 tags:
   - scientific-text
   - citation-intent
-  - citation-analysis
 datasets:
   - allenai/multicite
 widget:
@@ -19,9 +18,7 @@ widget:
 
 # SciBERT Citation-Background Classifier
 
-A [`allenai/scibert_scivocab_uncased`](https://huggingface.co/allenai/scibert_scivocab_uncased) model fine-tuned as a **binary classifier of citation intent**: does a citation sentence cite prior work as **background** (`BACKGROUND`) or for any other reason — method use, comparison, extension, motivation, etc. (`NOT_BACKGROUND`)?
-
-It was built during a large-scale **citation-fidelity** study as a filter for background citations. Note: in the paper's final pipeline this filtering was ultimately performed with Semantic Scholar's (S2ORC) built-in citation-intent labels rather than this model; it is released for transparency and reuse.
+[`allenai/scibert_scivocab_uncased`](https://huggingface.co/allenai/scibert_scivocab_uncased) fine-tuned as a binary classifier of citation intent: does a citation sentence cite prior work as background (`BACKGROUND`) or for any other reason (`NOT_BACKGROUND`)?
 
 ## Labels
 
@@ -42,46 +39,10 @@ clf("Deep learning has achieved remarkable success in many NLP tasks [1, 2].")
 
 ## Training
 
-- **Base model:** SciBERT (uncased, scivocab), fine-tuned with `BertForSequenceClassification` (single-label, 2 classes).
-- **Data:** binary labels derived from [MultiCite](https://github.com/allenai/multicite) (Lauscher et al., 2022): citation contexts with intent `background` → `BACKGROUND` (1), all other intents → `NOT_BACKGROUND` (0).
-- **Trained:** August 2023, `transformers` 4.32.0. The training script was not preserved, so exact hyperparameters are unknown.
+- **Base model:** SciBERT (uncased, scivocab; `BertForSequenceClassification`, single-label, 2 classes)
+- **Data:** citation contexts from [MultiCite](https://github.com/allenai/multicite) — intent `background` → `BACKGROUND`, all other intents → `NOT_BACKGROUND`
+- **Fine-tuned:** August 2023, `transformers` 4.32.0
 
 ## Evaluation
 
-Evaluated zero-shot on the **SciCite** test set (n = 1,861; label `background` vs. the rest) as a release sanity check — note SciCite is a *different* dataset/annotation scheme than the MultiCite-derived training data:
-
-| metric | value |
-|---------------------------------|-------|
-| Accuracy | 0.897 |
-| Majority-class baseline | 0.536 |
-| BACKGROUND precision | 0.888 |
-| BACKGROUND recall | 0.925 |
-| BACKGROUND F1 | 0.906 |
-
-No evaluation from the original training run survives; evaluate on your own data before critical use.
-
-## Limitations
-
-- Trained on citation **context sentences** from scientific papers; not suited to other text types.
-- The binary label collapses many distinct intents into `NOT_BACKGROUND`.
-- The label mapping was reconstructed from the project's dataset-construction notebook and verified empirically on SciCite (above).
-- This model was **not** used to produce the paper's final results (S2ORC intents were used instead).
-
-## Citation
-
-If you use this model, please cite the citation-fidelity paper (to appear — citation forthcoming). Related resources:
-
-```bibtex
-@inproceedings{lauscher2022multicite,
-  title={MultiCite: Modeling realistic citations requires moving beyond the single-sentence single-label setting},
-  author={Lauscher, Anne and Ko, Brandon and Kuehl, Bailey and Johnson, Sophie and Cohan, Arman and Jurgens, David and Lo, Kyle},
-  booktitle={NAACL},
-  year={2022}
-}
-@inproceedings{beltagy2019scibert,
-  title={SciBERT: A Pretrained Language Model for Scientific Text},
-  author={Beltagy, Iz and Lo, Kyle and Cohan, Arman},
-  booktitle={EMNLP},
-  year={2019}
-}
-```
+On the SciCite test set (n = 1,861; `background` vs. rest): accuracy **0.897**, `BACKGROUND` F1 **0.906**.