## Use
This model was created to classify texts according to their CEFR level on a scale from 0 to 5 (A1-C2).
## Training data and licenses
This model was fine-tuned (LoRA) on English texts from the
[UniversalCEFR](https://huggingface.co/UniversalCEFR) collection. Each source
dataset keeps its original license and copyright with its creators.
| Dataset | License | Reference |
|---|---|---|
| [cambridge_exams_en](https://huggingface.co/datasets/UniversalCEFR/cambridge_exams_en) | CC-BY-NC-SA-4.0 | (copy from dataset card) |
| [elg_cefr_en](https://huggingface.co/datasets/UniversalCEFR/elg_cefr_en) | CC-BY-NC-4.0 | Breuker (2023) |
| [cefr_sp_en](https://huggingface.co/datasets/UniversalCEFR/cefr_sp_en) | CC-BY-NC-SA-4.0 | (copy from dataset card) |
| [icle500_en](https://huggingface.co/datasets/UniversalCEFR/icle500_en) | CC-BY-NC-4.0 | (copy from dataset card) |
| [cefr_asag_en](https://huggingface.co/datasets/UniversalCEFR/cefr_asag_en) | CC-BY-NC-SA-4.0 | (copy from dataset card…
Regression model that scores English text on the CEFR scale. The output is a single number from 0 (A1) to 5 (C2); round it to get the level. Trained on all UniversalCEFR databases with Engish texts.
Training:
Split by document (80% train / 10% val / 10% test, fixed seed). The split happens before chunking, so pieces of the same text never end up in both train and test.
Chunk every text into pieces of up to 250 words.
Balance the levels. Every level is resampled to the median class size, so rare levels (A1, C2) are not ignored.
Augment. 30% of training chunks lose their punctuation and casing, to mimic YouTube auto-captions. Validation and test data are left untouched.
Train with LoRA (r=32, all linear layers, regression head fully trained). The best epoch by validation QWK is kept.
Metrics:
{'test_loss': 0.46324828267097473, 'test_qwk': 0.7706145974737674, 'test_adjacent_acc': 0.9694947569113441, 'test_macro_f1': 0.5760069348195234, 'test_exact_acc': 0.559580552907531, 'test_runti…
