Upload folder using huggingface_hub

README.md

@@ -0,0 +1,194 @@
+---
+license: mit
+tags:
+- text-classification
+- bert
+- orality
+- linguistics
+- rhetorical-analysis
+language:
+- en
+metrics:
+- f1
+- accuracy
+base_model:
+- google-bert/bert-base-uncased
+pipeline_tag: text-classification
+library_name: transformers
+datasets:
+- custom
+model-index:
+- name: bert-marker-type
+  results:
+  - task:
+      type: text-classification
+      name: Marker Type Classification
+    metrics:
+    - type: f1
+      value: 0.4486
+      name: F1 (macro)
+    - type: accuracy
+      value: 0.630
+      name: Accuracy
+---
+
+# Havelock Marker Type Classifier
+
+BERT-based classifier for **25 rhetorical marker types** on the oral–literate spectrum, grounded in Walter Ong's *Orality and Literacy* (1982).
+
+This is the mid-level of the Havelock span classification hierarchy. Given a text span identified as a rhetorical marker, the model classifies it into one of 25 functional types (e.g., `repetition`, `subordination`, `direct_address`, `hedging_qualification`).
+
+## Model Details
+
+| Property | Value |
+|----------|-------|
+| Base model | `bert-base-uncased` |
+| Architecture | `BertForSequenceClassification` |
+| Task | Multi-class classification (25 classes) |
+| Max sequence length | 128 tokens |
+| Best F1 (macro) | **0.4486** |
+| Best Accuracy | **0.630** |
+| Parameters | ~109M |
+
+## Usage
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+model_name = "HavelockAI/bert-marker-type"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+
+span = "whether or not the underlying assumptions hold true"
+inputs = tokenizer(span, return_tensors="pt", truncation=True, max_length=128)
+
+with torch.no_grad():
+    logits = model(**inputs).logits
+    pred = torch.argmax(logits, dim=1).item()
+
+print(f"Marker type: {model.config.id2label[pred]}")
+```
+
+## Label Taxonomy (25 types)
+
+The 25 types group the 72 fine-grained subtypes into functional families:
+
+| Oral Types | Literate Types |
+|------------|----------------|
+| `direct_address` | `subordination` |
+| `repetition` | `abstraction` |
+| `formulaic_phrases` | `hedging_qualification` |
+| `parallelism` | `analytical_distance` |
+| `parataxis` | `logical_connectives` |
+| `sound_patterns` | `textual_apparatus` |
+| `performance_markers` | `literate_feature` |
+| `concrete_situational` | `passive_agentless` |
+| `agonistic_framing` | |
+| `oral_feature` | |
+
+Legacy/low-support types also present in the label space: `agonistic`, `concrete`, `formulaic`, `hedging`, `logical_connective`, `passive`, `passive_constructions`. These have fewer than 10 test examples and the model does not reliably predict them.
+
+## Training
+
+### Data
+
+Span-level annotations from the same corpus as the category classifier. Each span carries a `marker_type` field. Only types with ≥15 examples in the full dataset are included; the rest are filtered out during label map construction.
+
+A stratified 80/20 train/test split was used (random seed 42). The test set contains 4,602 spans.
+
+### Hyperparameters
+
+| Parameter | Value |
+|-----------|-------|
+| Epochs | 3 |
+| Batch size | 8 |
+| Learning rate | 2e-5 |
+| Optimizer | AdamW |
+| LR schedule | Linear warmup (10% of total steps) |
+| Gradient clipping | 1.0 |
+| Loss | Cross-entropy |
+| Min examples per class | 15 |
+
+### Training Metrics
+
+| Epoch | Loss | Accuracy | F1 (macro) |
+|-------|------|----------|------------|
+| 1 | 1.9282 | 0.6043 | 0.4142 |
+| 2 | 1.1097 | 0.6215 | 0.4414 |
+| 3 | 0.7712 | 0.6297 | 0.4486 |
+
+Best checkpoint selected by F1 at epoch 3. Loss still declining steeply.
+
+### Test Set Classification Report
+
+<details><summary>Click to expand per-class precision/recall/F1/support</summary>
+```
+                       precision    recall  f1-score   support
+
+          abstraction      0.637     0.712     0.673       570
+            agonistic      0.000     0.000     0.000         7
+    agonistic_framing      0.902     0.698     0.787        53
+  analytical_distance      0.516     0.465     0.489       245
+             concrete      0.000     0.000     0.000         7
+ concrete_situational      0.471     0.467     0.469       225
+       direct_address      0.691     0.752     0.720       722
+            formulaic      0.000     0.000     0.000        10
+    formulaic_phrases      0.598     0.600     0.599       380
+              hedging      0.000     0.000     0.000        49
+hedging_qualification      0.477     0.588     0.527       194
+     literate_feature      0.690     0.703     0.696       111
+   logical_connective      0.000     0.000     0.000         5
+  logical_connectives      0.537     0.600     0.567       220
+         oral_feature      0.521     0.388     0.444        98
+          parallelism      0.786     0.805     0.795        41
+            parataxis      0.642     0.538     0.586       130
+              passive      0.000     0.000     0.000         4
+    passive_agentless      0.651     0.597     0.623       119
+passive_constructions      0.000     0.000     0.000         9
+  performance_markers      0.607     0.496     0.546       137
+           repetition      0.681     0.726     0.703       318
+       sound_patterns      0.693     0.591     0.638       149
+        subordination      0.663     0.689     0.676       586
+    textual_apparatus      0.708     0.648     0.676       213
+
+             accuracy                          0.630      4602
+            macro avg      0.459     0.443     0.449      4602
+         weighted avg      0.618     0.630     0.622      4602
+```
+
+</details>
+
+**Top performing types (F1 > 0.65):** `parallelism` (0.795), `agonistic_framing` (0.787), `direct_address` (0.720), `repetition` (0.703), `literate_feature` (0.696), `subordination` (0.676), `textual_apparatus` (0.676), `abstraction` (0.673).
+
+**Zero F1 types:** `agonistic`, `concrete`, `formulaic`, `hedging`, `logical_connective`, `passive`, `passive_constructions` — all have ≤10 test examples and appear to be legacy label variants superseded by more specific types.
+
+## Limitations
+
+- **Severely undertrained**: 3 epochs with loss at 0.77 and still falling sharply. This model would benefit substantially from more training.
+- **Label noise from legacy types**: 7 of 25 classes appear to be legacy/coarse variants that coexist with their refined replacements (e.g., `hedging` alongside `hedging_qualification`). This inflates the label space and depresses macro F1.
+- **Class imbalance**: `direct_address` has 722 test examples while `passive` has 4. Weighted F1 (0.622) is substantially higher than macro F1 (0.449), indicating the model performs better on common types.
+- **Span-level only**: Requires pre-extracted spans. Does not detect boundaries.
+- **128-token context window**: Longer spans are truncated.
+
+## Theoretical Background
+
+The type level captures functional groupings within the oral–literate framework. Oral types reflect Ong's characterization of oral discourse as additive (`parataxis`), aggregative (`formulaic_phrases`), redundant (`repetition`), agonistically toned (`agonistic_framing`), empathetic and participatory (`direct_address`), and close to the human lifeworld (`concrete_situational`). Literate types capture the analytic (`abstraction`, `subordination`), distanced (`analytical_distance`, `passive_agentless`), and self-referential (`textual_apparatus`) qualities of written discourse.
+
+## Citation
+```bibtex
+@misc{havelock2026type,
+  title={Havelock Marker Type Classifier},
+  author={Havelock AI},
+  year={2026},
+  url={https://huggingface.co/HavelockAI/bert-marker-type}
+}
+```
+
+## References
+
+- Ong, Walter J. *Orality and Literacy: The Technologizing of the Word*. Routledge, 1982.
+
+---
+
+*Trained: February 2026*
+*Model version: da931b4a · Trained: February 2026*

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d9f0f0f5c783d07236ac2b0fc8982b1daa31efb9abfd9db72d011921d6b6c1f8
+oid sha256:18737f307d25ae24953445bd387589b30067adc7557d0506b3311953b9a8bd6f
 size 438029372