README.md

---
base_model: Delta-Vector/Control-8B-V1.1
library_name: transformers
model_name: controlkto
tags:
- generated_from_trainer
- axolotl
- trl
- kto
licence: license
---

# Model Card for controlkto

This model is a fine-tuned version of [Delta-Vector/Control-8B-V1.1](https://huggingface.co/Delta-Vector/Control-8B-V1.1).
It has been trained using [TRL](https://github.com/huggingface/trl).

## Quick start

```python
from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="jeiku/controlkto", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])
```

## Training procedure

[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/welchjacob254/controlkto/runs/r5p8sb5l)

This model was trained with KTO, a method introduced in [KTO: Model Alignment as Prospect Theoretic Optimization](https://huggingface.co/papers/2402.01306).

### Framework versions

- TRL: 0.12.1
- Transformers: 4.47.0
- Pytorch: 2.3.1+cu121
- Datasets: 3.1.0
- Tokenizers: 0.21.0

## Citations

Cite KTO as:

```bibtex
@article{ethayarajh2024kto,
    title        = {{KTO: Model Alignment as Prospect Theoretic Optimization}},
    author       = {Kawin Ethayarajh and Winnie Xu and Niklas Muennighoff and Dan Jurafsky and Douwe Kiela},
    year         = 2024,
    eprint       = {arXiv:2402.01306},
}
```

Cite TRL as:
    
```bibtex
@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouédec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}
```