Fine-tune bert-medium on high-quality NLI subset

final/README.md

@@ -0,0 +1,70 @@
+---
+language: en
+license: apache-2.0
+tags:
+- text-classification
+- zero-shot-classification
+- nli
+datasets:
+- Pankaj8922/nli-high-quality-balanced
+base_model: prajjwal1/bert-medium
+metrics:
+- accuracy
+- f1
+model-index:
+- name: bert-medium-nli
+  results:
+  - task:
+      type: text-classification
+      name: Natural Language Inference
+    dataset:
+      name: Pankaj8922/nli-high-quality-balanced
+      type: Pankaj8922/nli-high-quality-balanced
+    metrics:
+    - type: accuracy
+      value: 0.8373
+      name: Test Accuracy
+    - type: f1
+      value: 0.8372
+      name: Test F1 (macro)
+---
+
+# bert-medium-nli
+
+Fine-tuned [`prajjwal1/bert-medium`](https://huggingface.co/prajjwal1/bert-medium) for natural language
+inference (entailment / neutral / contradiction), intended for use as a zero-shot
+text classification model via the entailment trick (hypothesis = "This text is
+about {label}.").
+
+Trained on [`Pankaj8922/nli-high-quality-balanced`](https://huggingface.co/datasets/Pankaj8922/nli-high-quality-balanced), a
+combined and filtered subset of MNLI, SNLI, FEVER-NLI, and ANLI:
+annotator-agreement filtered, deduplicated, teacher-confidence filtered,
+hypothesis-only artifact filtered, and class-balanced.
+
+## Results
+
+| Split | Accuracy | F1 (macro) | Precision (macro) | Recall (macro) |
+|---|---|---|---|---|
+| Validation | 0.8389 | 0.8389 | 0.8389 | 0.8389 |
+| Test | 0.8373 | 0.8372 | 0.8372 | 0.8373 |
+
+## Training details
+
+- Base model: `prajjwal1/bert-medium`
+- Epochs: 3
+- Batch size: 64 (train), 128 (eval)
+- Learning rate: 5e-05
+- Max sequence length: 256
+
+## Labels
+
+- 0: entailment
+- 1: neutral
+- 2: contradiction
+
+## Intended use / limitations
+
+This is a small (~41M parameter) model, so its ceiling on zero-shot performance
+against novel, unseen label sets is lower than larger NLI-tuned checkpoints
+(e.g. DeBERTa-v3-base or -large variants). Best suited for fast inference or
+resource-constrained settings rather than maximum accuracy.