---
license: unknown
language:
- en
tags:
- wine
- clm
- llm
pipeline_tag: text2text-generation
widget:
- text: "describe: Pinot Noir Russian River Valley US"
inference:
  parameters:
    max_new_tokens: 128
    num_beams: 5
    no_repeat_ngram_size: 2
---

# Wineberto T5

T5 model fine tuned on the winemag reviews dataset to generate wine descriptions using text to text generation. There is some correlation between the wine label and the description based on whats provided in the inputs such as wine variety, country and region.

## Model description


## How to use

You can use this model directly like so. Prefix the wine label with the string `describe: `

```python
>>> from transformers import pipeline
>>> seq = pipeline('text2text-generation', model='panigrah/winberto-t5-s2s')
>>> seq(["describe: Pinot Noir Russian River Valley US", "describe: Pinot Noir Willamette Valley Oregon"],  max_new_tokens=128, num_beams=5, no_repeat_ngram_size=2, early_stopping=True)

[{'generated_text': "This is a light, fruity Pinot Noir, with aromas of red cherry, cranberry and pomegranate. It's light in body and light on the palate, offering flavors of cherry and raspberry. The finish is clean and clean."},
 {'generated_text': "This is a ripe, full-bodied Pinot Noir. It's full bodied and supple in feel, with flavors of cherry, raspberry, cola and spice. The tannins are firm and firm, and the wine is ready to drink."}]

```
## Training data

The T5-small model was fine tuned with 150K wine review descriptions and with a ROUGE metric. 
Learning rate was kept high to make the model converge faster - higher LRs work better for T5 based on [this discussion](https://discuss.huggingface.co/t/t5-finetuning-tips/684/34)
Used a [ROUGE](https://stats.stackexchange.com/questions/301626/interpreting-rouge-scores) metric to measure performance - highest ROUGE on stopping training had ROUGE1 of around 25 and ROUGE2 of about 15. 
Will continue trying to tune training parameters to see if we can get better results.