Spaces:
Running
Model Card 路 ModelOps Research Lab baselines
Intended purpose
Research demonstration of reproducible, browser-trained multilingual academic-sentence classification. Labels: background, objective, method, result, limitation, implication.
Model architectures
Not neural networks. Four transparent lexical baselines: multinomial Naive Bayes, word TF-IDF class centroid, character trigram TF-IDF class centroid and a fixed 55/45 centroid hybrid. No pretrained weights and no API inference.
Training / evaluation
Training on the 842-row original fixed synthetic-curated training set (EN / ES / PT). Held-out assessments use 130 controlled IID examples and 54 fixed novel-pattern synthetic examples. Benchmark code is published in engine.js and reproduced by benchmark_cli.js. The release metrics are in evaluation_results.json.
Scores
| Model | IID Accuracy | Challenge Accuracy | Challenge Macro F1 |
|---|---|---|---|
| NB | 100.0% | 74.1% | 73.6% |
| Word centroid | 98.5% | 72.2% | 70.6% |
| Char centroid | 95.4% | 70.4% | 68.4% |
| Hybrid | 97.7% | 77.8% | 76.2% |
No external human-labeled validation has been performed for these browser models. A separate archived Python classifier was once smoke-tested on PubMed; those figures are contextual only and not directly comparable.
Known limitations and risks
Controlled synthetic data, likely recurring templates, limited domain scope, demographic/fairness relevance not assessed, no calibrated probabilities, unreliable transfer to real scholarly prose. Prohibited for grading, academic-integrity judgments, student assessment, admissions, employment selection or publication decisions.
License
MIT code; CC BY 4.0 training/evaluation dataset (see DATA_LICENSE.md).