mskane968/cars-scibert

CARS SciBERT Rhetorical Move Classifier

Fine-tuned allenai/scibert_scivocab_uncased to predict the three CARS rhetorical moves (Establishing Territory, Establishing Niche, Presenting Present Work) from individual sentences collected out of academic article introductions.

Training Details

• Dataset: /Users/megankane/Documents/CARS_Classifier/coded_CARS_sentences.csv (Establishing Territory 823, Establishing Niche 509, Presenting Present Work 442)
• Split: 80% train / 20% eval stratified by label
• Hyperparameters: max length 192, batch sizes 8/16 (train/eval), learning rate 2e-5, weight decay 0.01, warmup ratio 0.06, epochs 2, gradient accumulation 1, TrainingArguments(eval_strategy="epoch", save_strategy="epoch")
• Hardware: CPU fine-tuning via Hugging Face Trainer

Metrics (eval split)

Metric	Value
Accuracy	0.608
Macro F1	0.577
Establishing Niche F1	0.413
Establishing Territory F1	0.699
Presenting Present Work F1	0.617

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "mskane968/cars-scibert"  # replace with actual repo name after upload
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

texts = [
    "Previous studies have highlighted the importance of translanguaging in multilingual classrooms.",
]
inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True)
outputs = model(**inputs)
pred_ids = outputs.logits.argmax(dim=-1)
label_map = ['Establishing Niche', 'Establishing Territory', 'Presenting Present Work']
print([label_map[i] for i in pred_ids])

Files

• config.json, pytorch_model.bin – fine-tuned SciBERT weights
• tokenizer.json, tokenizer_config.json, vocab.txt, special_tokens_map.json – tokenizer assets
• label_encoder.npy – label order for downstream consumers (optional)
• training_args.bin – Hugging Face Trainer arguments for reproducibility

License

Released under the MIT License (see LICENSE).

Limitations & Intended Use

• Model was trained on a moderate (~1.8k sentences) academic corpus and may not generalize to other writing styles.
• Establishing Niche recall is still limited; predictions should be reviewed before automation-critical decisions.
• Not evaluated for fairness/bias; do not apply to sensitive content without further analysis.