---
license: apache-2.0
base_model: FacebookAI/roberta-large
datasets:
- google-research-datasets/go_emotions
language:
- en
tags:
- goemotions
- emotion-classification
- multi-label-classification
- roberta
- focal-loss
- threshold-optimization
metrics:
- f1
pipeline_tag: text-classification
library_name: transformers
---

# GoEmotions RoBERTa Large Focal SOTA

RoBERTa-large multi-label classifier fine-tuned on the public GoEmotions
`simplified` split. The model predicts 28 labels: 27 fine-grained emotions plus
`neutral`.

## Results

The selected policy is coordinate threshold search optimized on validation
macro-F1.

| Split | Macro F1 | Micro F1 | Samples F1 | Hamming Loss |
| --- | ---: | ---: | ---: | ---: |
| Validation | 0.565864 | 0.596637 | 0.605125 | 0.034424 |
| Test | 0.533020 | 0.576651 | 0.585942 | 0.035820 |

The per-label threshold candidate reached test macro-F1 `0.534994`, but the
validation-selected policy is the coordinate threshold policy above.

## Training

- Base model: `FacebookAI/roberta-large`
- Dataset: `google-research-datasets/go_emotions`, `simplified`
- Loss: focal loss, alpha `0.38`, gamma `2.8`
- Epochs: `4`
- Learning rate: `1e-5`
- Batch: `2`, gradient accumulation `16`
- Mixed precision: disabled for stability on Kaggle T4
- Seed: `42`

Training provenance: Kaggle kernel
`kevin250304/goemotions-roberta-large-focal-sweep`.

## Usage

```python
import json
import numpy as np
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo_id = "AliceYin/goemotions-roberta-large-focal-sota"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

labels = json.load(open("labels.json"))["label_names"]
thresholds = json.load(open("thresholds.json"))["coordinate"]
threshold_array = np.array([thresholds[label] for label in labels])

text = "I can't believe this worked so well."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    probs = torch.sigmoid(model(**inputs).logits).cpu().numpy()[0]

predicted = [
    label for label, prob, threshold in zip(labels, probs, threshold_array)
    if prob >= threshold
]
print(predicted)
```

For remote loading of thresholds and labels, download `labels.json` and
`thresholds.json` from this repository alongside the model files.

## Files

- `model.safetensors`: fine-tuned RoBERTa-large weights
- `config.json`, `tokenizer.json`, `tokenizer_config.json`: Transformers files
- `thresholds.json`: fixed, global, per-label, and coordinate thresholds
- `labels.json`: label names and neutral index
- `metrics.json`: full validation/test metrics
- `kaggle-run.log`: training log

## Limitations

This model is trained on Reddit comments and inherits the dataset's domain,
annotation, and class imbalance limitations. Validate before using in
high-stakes settings.