File size: 2,778 Bytes
5147a82 1f43cef 5147a82 1f43cef 5147a82 1f43cef 5147a82 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 | ---
language: en
license: other
library_name: transformers
tags:
- text-classification
- distilbert
- podcast
- ad-detection
- skipr
datasets:
- custom
base_model: distilbert-base-uncased
pipeline_tag: text-classification
---
# Skipr Ad Classifier
Fine-tuned DistilBERT model that detects sponsor/ad segments in YouTube podcast transcript text.
**Also available:** [dkayaaaa/ad-classifier-quantised](https://huggingface.co/dkayaaaa/ad-classifier-quantised) — INT8 ONNX version (~64 MB) for faster, lighter inference with ONNX Runtime.
## Model description
This model classifies a short transcript window as either an ad/sponsor segment or normal podcast content. It was trained as part of the [Skipr](https://github.com/YOUR_ORG/skippy-model-training) pipeline for skipping sponsor segments in YouTube podcasts.
- **Architecture:** `distilbert-base-uncased`
- **Task:** Binary sequence classification
- **Labels:**
- `0` — not an ad segment
- `1` — ad/sponsor segment
- **Max sequence length:** 512 tokens
- **Training:** 3 epochs, fine-tuned from `distilbert-base-uncased`
## Intended use
Use this model to classify transcript windows (typically ~20 caption snippets) as ad vs non-ad content. It is designed for use in the Skipr browser extension and related inference services.
**Out of scope:**
- General sentiment or topic classification
- Non-English text (trained on English podcast transcripts)
- Full-video classification without segmentation
## Training data
## Training data
The model was trained on a mix of real and synthetic transcript windows:
- **Base set (~800 samples):** weak-labeled YouTube podcast segments
- Positive: segments matching sponsor keywords/brands
- Negative: normal podcast content
- **Augmented set (~1,200 samples):** synthetic variants generated with Llama 8B via Ollama, preserving the original label. Synthetic data generated through strategies; paraphrase, new scenario, style shift, fragment, vocabulary shift
Original labels are heuristic — the model learns from keyword-labeled examples. Synthetic data increases linguistic diversity but inherits the same label assumptions.
## Usage
### Transformers (Python)
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
repo = "dkayaaaa/ad-classifier"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)
text = "this episode is brought to you by our friends at..."
inputs = tokenizer(
text,
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=512,
)
with torch.no_grad():
logits = model(**inputs).logits
prediction = logits.argmax(dim=-1).item()
print("ad" if prediction == 1 else "not ad") |