File size: 8,661 Bytes
33de8ff |
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 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 |
<h1 align="center">
<u>C</u>ritique-out-<u>Loud</u> Reward Models (CLoud)
</h1>
<p align="center">
<img src="assets/CLoud.gif" alt="CLoud"/>
</p>
<p align="center">
| <a href="https://arxiv.org/abs/2408.11791"><b>Paper</b></a> | <a href="https://x.com/ZackAnkner/status/1826607200376336478"> <b>Tweet</b> </a> |
</p>
---
## Introduction
<u>C</u>ritique-out-<u>Loud</u> reward models are reward models that can reason explicitly about the quality of an input through producing Chain-of-Thought like critiques of an input before predicting a reward.
In classic reward model training, the reward model is trained as a reward head initialized on top of the base LLM.
Without LM capabilities, classic reward models act as encoders and must predict rewards within a single forward pass through the model, meaning reasoning must happen implicitly.
In contrast, CLoud reward models are trained to both produce explicit reasoning about quality and to score based on these critique reasoning traces.
CLoud reward models lead to large gains for pairwise preference modeling on RewardBench, and also lead to large gains in win rate when used as the scoring model in Best-of-N sampling on ArenaHard.
## Todo
- [x] Release models and inference examples
- [ ] Post example training run logs
- [ ] Add ArenaHard evaluation code
- [ ] Add VLLM support for inference
## Table of Contents
- [Introduction](#introduction)
- [Todo](#todo)
- [Table of Contents](#table-of-contents)
- [Setup](#setup)
- [Model Weights](#model-weights)
- [Inference](#inference)
- [Dataset](#dataset)
- [Training](#training)
- [CLoud Training](#cloud-training)
- [Classic Training](#classic-training)
- [Evaluation](#evaluation)
- [Citation](#citation)
## Setup
```bash
git clone https://github.com/zankner/CLoud
cd CLoud
pip install -e .
```
Optional: base docker image used during development `mosaicml/pytorch:2.3.0_cu121-python3.11-ubuntu20.04`
## Model Weights
| Base Model | RM Type | Hugging Face Repo |
| ---------- | --------------- |--------------------------------------------------------------------- |
| Llama3-8B | Classic | [ankner/Llama3-8B-Classic-RM](https://huggingface.co/ankner/Llama3-8B-Classic-RM) |
| Llama3-8B | CLoud | [ankner/Llama3-8B-CLoud-RM](https://huggingface.co/ankner/Llama3-8B-CLoud-RM) |
| Llama3-70B | Classic | [ankner/Llama3-70B-Classic-RM](https://huggingface.co/ankner/Llama3-70B-Classic-RM) |
| Llama3-70B | CLoud | [ankner/Llama3-70B-CLoud-RM](https://huggingface.co/ankner/Llama3-70B-CLoud-RM) |
## Inference
We provide a gradio demo which can be run as follows: `gradio cloud/demo.py`. By default this will demo `ankner/Llama3-8B-CLoud-RM`, but you can change the model loaded in the script.
If you want to perform inference on your own data, please refer to the following example:
```python
from cloud.model import CLoudRewardModel
from transformers import AutoTokenizer
model_name = "ankner/Llama3-8B-Cloud-RM" # Replace with RM trained with this repo
model = CLoudRewardModel.from_pretrained(model_name, device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="left")
user_prompt = [
"Write me a story",
"What is the capital of the moon?"
]
assistant_response = [
"No I don't want to do that.",
"Since the moon is made out of cheese, the capital is mozzerella."
]
rewards, critiques = model.predict_reward(user_prompt, assistant_response, tokenizer)
for reward, critique in zip(rewards, critiques):
print("Critique:")
print(critique)
print("Reward:")
print(reward)
print("=" * 100)
```
## Dataset
We provide code to reconstruct the datasets used in the paper.
There are two datasets to build for training, one with oracle critiques meant to simmulate human feedback and one with self-generated critiques.
To build the oracle critique dataset run:
```bash
python cloud/data/build_official_ultra_llama.py --mode oracle
```
To build the self-generated critique dataset run:
```bash
python cloud/data/build_official_ultra_llama.py --mode self-gen --model-size {model-size}
```
where ```{model-size}``` is the size of the model you are using (e.g. 8b, 70b).
<details>
<summary>Build your own dataset from scratch</summary>
1. <b>Build prompts</b> - You can use any dataset you like as long as it has ```prompt``` and ```id``` columns. If you would like to build prompts from UltraFeedback and UltraInteract as we do in the paper run:
```bash
python cloud/data/build_ultra_prompts.py --save-name {name-to-save-as}
```
2. <b>Build chosen / rejected responses</b>
```bash
python cloud/data/build_judgements.py --gen-model {model-generating-responses} --judge-model {model-judging-responses} --base-dataset {path-to-prompt-dataset} --save-name {name-to-save-as}
```
The above command requires a hosted generating and judging model. To host the models using vllm run:
```bash
python -m vllm.entrypoints.openai.api_server --model {path-to-gen/judge-model} --dtype bfloat16 --tensor-parallel-size {num-gpus} --port {8000 for gen and 8001 for judge}
```
3. <b>Build critiques</b>
```bash
python cloud/data/generate_oracle_critiques.py --judge-model {model-generating-critiques} --base-dataset {path-to-responses-dataset} --save-name {name-to-save-as}
```
Again, this command assumes a hosted critique model. To host the critique model you can use the above vllm command (This time just use port 8000 for the judge model).
</details>
## Training
Before training, you must run the [setup script](#setup) and build the [datasets](#dataset).
The training configs are located in the ```cloud/train/configs/``` folder.
We have already set the optimal hyperparameters that we found for each model as reported in the paper.
The only parameter that needs to be set is the ```variables.micro_batch_size``` parameter, in accordance with your GPU memory.
If you want to log the training runs, uncomment the ```loggers``` section in the config and fill in your wandb settings.
Checkpoints will be saved throughout training to the ```save_folder``` parameter, which is ```ckpts/${variables.run_name}``` by default. The final checkpoint will contain a folder ```hf``` where the huggingface model is saved.
> **Warning**: The below training scripts for both CLoud and Classic prefill the dataset names to be the datasets we release. If you would like to train on your own dataset, you will need to follow the directions to build said dataset in the [dataset section](#dataset) and change the ```variables.dataset_path``` parameter in the training configs.
### CLoud Training
1. The first step is to finetune the base model to produce critiques:
```bash
composer -n {num_gpus} cloud/train/train.py cloud/train/configs/{model_size}_critique_sft.yaml
```
Replace ```{model_size}``` with the size of the model you are training (e.g. 8b, 70b).
2. (Optional if you want to use the self-generated data we release) After the critique SFT model is trained, you need to regenerate the dataset with the critiques.
To do so, you first need to serve the critique SFT model. To do so locally using vllm run:
```bash
python -m vllm.entrypoints.openai.api_server --model {path-to-critique-sft-model} --dtype bfloat16 --tensor-parallel-size {num-gpus}
```
Then run the data building script:
```bash
python cloud/data/generate_self_critiques.py --model {path-to-critique-sft-model} --base-dataset {path-to-base-dataset} --upload-name {path-to-save-dataset}
```
3. After building the self-generated dataset, we can train the CLoud model:
```bash
composer -n {num_gpus} cloud/train/train.py cloud/train/configs/{model_size}_cloud.yaml
```
### Classic Training
To train a classic reward model, you can use the following command:
```bash
composer -n {num_gpus} cloud/train/train.py cloud/train/configs/{model_size}_classic.yaml
```
## Evaluation
To run evaluation for a given benchmark run the following command:
```bash
python cloud/eval/eval.py --model-path {path-to-model} --benchmark {benchmark-name}
```
Currently, we only support the RewardBench benchmark.
## Citation
If you found our work useful please consider citing it:
```bibtex
@misc{ankner2024critiqueoutloudrewardmodels,
title={Critique-out-Loud Reward Models},
author={Zachary Ankner and Mansheej Paul and Brandon Cui and Jonathan D. Chang and Prithviraj Ammanabrolu},
year={2024},
eprint={2408.11791},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2408.11791},
}
``` |