---
base_model: roberta-base
library_name: transformers
license: mit
pipeline_tag: text-classification
metrics:
- accuracy
tags:
- generated_from_trainer
model-index:
- name: roberta-sentence-classifier
  results: []
---

# roberta-sentence-classifier

This model is a fine-tuned version of [roberta-base](https://huggingface.co/roberta-base) presented in the paper **[Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction](https://huggingface.co/papers/2606.28186)**.

It serves as the sentence-level cognitive episode tagger in the **Epi2Diff** (Episode to Difficulty) framework. It maps Large Reasoning Model (LRM) reasoning traces into cognitively grounded episode sequences to support interpretable modeling of human item difficulty.

- **Repository:** [c-steve-wang/Epi2Diff](https://github.com/c-steve-wang/Epi2Diff)
- **Paper:** [Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction](https://huggingface.co/papers/2606.28186)

## Model description

The model classifies sentence-level reasoning units into 8 problem-solving cognitive episode states:
- `Read`
- `Analyze`
- `Plan`
- `Implement`
- `Explore`
- `Verify`
- `Monitor`
- `Answer`

These classified sequences are then used by the Epi2Diff framework to extract compact episode-dynamic process features for downstream item difficulty prediction.

## Intended uses & limitations

You can use this model to segment and tag raw reasoning traces into functional problem-solving states to evaluate reasoning behaviors, perform interpretability studies, or support downstream educational measurement tasks.

## Training and evaluation data

The model was fine-tuned on annotated reasoning trace sentences derived from datasets such as SAT Math, SAT Reading & Writing, Cambridge, and USMLE.

It achieves the following results on the evaluation set:
- Loss: 0.6266
- Accuracy: 0.7990
- Macro F1: 0.7614
- Micro F1: 0.7990
- Qwk: 0.6588

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- num_epochs: 4
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step   | Validation Loss | Accuracy | Macro F1 | Micro F1 | Qwk    |
|:-------------:|:-----:|:------:|:---------------:|:--------:|:--------:|:--------:|:------:|
| 0.6267        | 1.0   | 27540  | 0.6108          | 0.7818   | 0.7364   | 0.7818   | 0.6352 |
| 0.5539        | 2.0   | 55080  | 0.5939          | 0.7911   | 0.7498   | 0.7911   | 0.6428 |
| 0.475         | 3.0   | 82620  | 0.6021          | 0.7977   | 0.7592   | 0.7977   | 0.6599 |
| 0.4204        | 4.0   | 110160 | 0.6266          | 0.7990   | 0.7614   | 0.7990   | 0.6588 |


### Framework versions

- Transformers 4.57.3
- Pytorch 2.9.0+cu126
- Datasets 4.0.0
- Tokenizers 0.22.1

## Citation

If you use this model, please cite:

```bibtex
@misc{wang2026epi2diff,
  title = {Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction},
  author = {Wang, Chenguang and Li, Ming and Zeng, Xinyue and Li, Zhuochun and Jiao, Hong and Zhou, Tianyi and Zhou, Dawei},
  year = {2026}
}
```