---
license: apache-2.0
datasets:
- your-username/cuebench
language:
- en
tags:
- contextual-prediction
- autonomous-driving
- reasoning
- multi-label
metrics:
- f1
- precision
- recall
model-index:
- name: <Model Name> on CUEBench
  results:
    - task:
        type: contextual-prediction
        name: Contextual Entity Prediction
      dataset:
        name: CUEBench
        type: your-username/cuebench
        config: default
        split: test
      metrics:
        - name: F1
          type: f1
          value: 0.74
        - name: Precision
          type: precision
          value: 0.76
        - name: Recall
          type: recall
          value: 0.71
---

# <Model Name> on CUEBench 

This model was evaluated on [**CUEBench**](https://huggingface.co/datasets/your-username/cuebench), a benchmark for **contextual entity prediction** in real-world autonomous driving scenes. The task requires predicting **unobserved or occluded entities** based on observed scene context.

## Description

- **Task**: Multi-label prediction of unobserved classes based on scene context.
- **Input**: List of observed classes (e.g., `["Car", "Pedestrian"]`)
- **Output**: Predicted target classes (e.g., `["PickupTruck", "Bus"]`)

## Training Data

The model was fine-tuned on the training split of CUEBench:
- ~8,000 scene-level examples
- Each example includes an `observed_classes` field and a `target_classes` label

## Training Procedure

- Base model: `bert-base-uncased` (or any other model)
- Input tokenized as string: `"Observed: Car, Pedestrian"`
- Output trained as multi-label classification over predefined entity set
- Optimizer: AdamW, learning rate: 5e-5
- Epochs: 5

## Evaluation

Evaluated on CUEBench test set using:
- **F1**: 0.74
- **Precision**: 0.76
- **Recall**: 0.71

Evaluation script:
```python
from datasets import load_dataset, load_metric

dataset = load_dataset("your-username/cuebench", split="test")
metric = load_metric("your-username/cuebench-metric")

predictions = [...]  # List of predicted sets
references = [json.loads(x["target_classes"]) for x in dataset]

result = metric.compute(predictions=predictions, references=references)