---
license: apache-2.0
datasets:
- kelm
language:
- en
library_name: peft
pipeline_tag: text2text-generation
---

This is a version of `flan-t5-xl` fine-tuned on the [KELM Corpus](https://github.com/google-research-datasets/KELM-corpus) to take in sentences and output triplets of the form `subject-relation-object` to be used for knowledge graph generation.

The model uses custom tokens to delimit triplets: 
```
special_tokens = ['<triplet>', '</triplet>', '<relation>', '<object>']
tokenizer.add_tokens(special_tokens)
```

You can use it like this:
```
model = model.to(device)
model.eval()

new_input = "Hugging Face, Inc. is an American company that develops tools for building applications using machine learning.",
inputs = tokenizer(new_input, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(input_ids=inputs["input_ids"].to("cuda"))
    print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=False)[0])
```
Output: `<pad><triplet> Hugging Face <relation> instance of <object> Business </triplet></s>`

This model still isn't perfect, and may make mistakes! I'm working on fine-tuning it for longer and on a more diverse set of data.