Cornell-AGI/Ultrafeedback-Llama-3-Armo-iter_3

---
dataset_info:
features:
- name: response_0
dtype: string
- name: response_1
dtype: string
- name: response_2
dtype: string
- name: response_3
dtype: string
- name: response_4
dtype: string
- name: prompt_id
dtype: string
- name: prompt
dtype: string
- name: llama_prompt
dtype: string
- name: llama_prompt_tokens
sequence: int64
- name: response_0_reward
dtype: float64
- name: response_1_reward
dtype: float64
- name: response_2_reward
dtype: float64
- name: response_3_reward
dtype: float64
- name: response_4_reward
dtype: float64
- name: chosen
dtype: string
- name: chosen_reward
dtype: float64
- name: llama_chosen
dtype: string
- name: llama_chosen_tokens
sequence: int64
- name: reject
dtype: string
- name: reject_reward
dtype: float64
- name: llama_reject
dtype: string
- name: llama_reject_tokens
sequence: int64
- name: chosen_logprob
dtype: float64
- name: reject_logprob
dtype: float64
splits:
- name: train_prefs
num_bytes: 2288598631
num_examples: 43188
- name: test_prefs
num_bytes: 75595359
num_examples: 1435
download_size: 598464955
dataset_size: 2364193990
configs:
- config_name: default
data_files:
- split: train_prefs
path: data/train_prefs-*
- split: test_prefs
path: data/test_prefs-*
---

# Dataset Card for Ultrafeedback-Llama-3-Armo-iter_3

This dataset was used to train [REBEL-Llama-3-Armo-iter_3](https://huggingface.co/Cornell-AGI/REBEL-Llama-3-Armo-iter_3).

We generate 5 responses using [REBEL-Llama-3-Armo-iter_2](https://huggingface.co/Cornell-AGI/REBEL-Llama-3-Armo-iter_2) and collect the rewards with [ArmoRM-Llama3-8B-v0.1](https://huggingface.co/RLHFlow/ArmoRM-Llama3-8B-v0.1). The best response in terms of reward is selected as chosen while the worst is selected as reject.

The 'chosen_logprob' and 'reject_logprob' are calculated based on [REBEL-Llama-3-Armo-iter_2](https://huggingface.co/Cornell-AGI/REBEL-Llama-3-Armo-iter_2). Note that these values may differ based on the cuda version and GPU configurations. Please consider recompute these values for your own experiments.

### Evaluations

| Model | AlpacaEval 2.0<br>LC Win Rate | AlpacaEval 2.0<br>Win Rate | MT-Bench<br>Average | MMLU<br>(5-shot) | GSM8K<br>(5-shot) |
| :--------: | :--------: | :--------: | :--------: | :--------: | :--------: |
| REBEL-OpenChat-3.5| 17.3 | 12.8 | 8.06 | 63.7 | 68.8 |
| REBEL-Llama-3 | 30.1 | 32.6 | 8.16 | 65.8 | 75.6 |
| REBEL-Llama-3-epoch_2| 31.3 | 34.2 | 7.83 | 65.4 | 75.4 |
| REBEL-Llama-3-Armo-iter_1| 48.3 | 41.8 | 8.13 | 66.3 | 75.8 |
| REBEL-Llama-3-Armo-iter_2| 50.0 | 48.5 | 8.07 | 65.9 | 75.4 |
| REBEL-Llama-3-Armo-iter_3| 49.7 | 48.1 | 8.01 | 66.0 | 75.7 |

## Citation
Please cite our paper if you use this dataset in your own work:
```
@misc{gao2024rebel,
title={REBEL: Reinforcement Learning via Regressing Relative Rewards},
author={Zhaolin Gao and Jonathan D. Chang and Wenhao Zhan and Owen Oertell and Gokul Swamy and Kianté Brantley and Thorsten Joachims and J. Andrew Bagnell and Jason D. Lee and Wen Sun},
year={2024},
eprint={2404.16767},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```