---
language:
- en
license: apache-2.0
tags:
- text-classification
- topic-classification
- deberta
- the-founder
datasets:
- fancyzhx/ag_news
metrics:
- accuracy
base_model: microsoft/deberta-v3-base
pipeline_tag: text-classification
model-index:
- name: founder-deberta-agnews-v1
  results:
  - task:
      type: text-classification
      name: Text Classification
    dataset:
      name: AG News
      type: fancyzhx/ag_news
      split: test
    metrics:
    - type: accuracy
      value: 0.2500
      name: Accuracy
    - type: loss
      value: nan
      name: Eval Loss
---

# founder-deberta-agnews-v1

Fine-tuned by **The Founder** — an autonomous ML orchestration superagent.

## Model Description

Fine-tuned `microsoft/deberta-v3-base` on `fancyzhx/ag_news` for 4-class topic classification (World, Sports, Business, Sci/Tech). Orchestrated end-to-end by **The Founder** using Tesla T4 (Kaggle), Weights & Biases, and HuggingFace Hub.

## Model Details

| Property | Value |
|---|---|
| Base model | microsoft/deberta-v3-base |
| Fine-tuned on | fancyzhx/ag_news |
| Task | 4-class topic classification |
| Labels | World, Sports, Business, Sci/Tech |
| Epochs | 3 |
| Batch size | 16 |
| Learning rate | 1e-05 |
| GPU | Tesla T4 (Kaggle) |
| Train loss | 0.6909 |
| Test loss | nan |
| Test accuracy | 0.2500 |
| Duration | 42.9 min |

## How to Get Started

```python
from transformers import pipeline
clf = pipeline("text-classification", model="zanesmit29/founder-deberta-agnews-v1")
clf("Apple announces new MacBook with M4 chip")
```

## Uses

### Direct Use
Topic classification of English news articles across 4 categories: World, Sports, Business, and Science/Technology.

### Out-of-Scope Use
Not suitable for non-English text, fine-grained news categorisation, or documents outside the news domain.

## Training Details

### Data
`fancyzhx/ag_news` — 120k train articles. 10% held out as validation (seed=42). 7.6k held-out test set used for final evaluation.

### Hyperparameters

| Hyperparameter | Value |
|---|---|
| Learning rate | 1e-05 |
| Batch size | 16 |
| Epochs | 3 |
| Optimizer | AdamW |
| LR scheduler | Linear with warmup (10%) |
| Max sequence length | 128 |
| fp16 | true |

## Results

| Metric | Value |
|---|---|
| Train loss | 0.6909 |
| Test loss | nan |
| Test accuracy | 0.2500 |
| Duration | 42.9 min |
| Hypothesis (>= 0.945) | REJECTED |

## Experiment Tracking

[W&B Run](https://wandb.ai/zane-ml/the-founder/runs/sip1yk0c)

## Infrastructure

| Component | Tool |
|---|---|
| Compute | Tesla T4 (Kaggle) |
| Experiment tracking | Weights & Biases |
| Artifact storage | HuggingFace Hub |
| Orchestration | The Founder |