Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture. Basically ChatGPT but with PaLM
| Date | Stars |
|---|---|
| 2026-07-24 | 7866 |
| 2026-07-25 | 7866 |
| 2026-07-28 | 7864 |
| 2026-07-30 | 7864 |
| 2026-07-31 | 7863 |
| 2026-08-02 | 7864 |
| 2026-08-03 | 7865 |
| 2026-08-06 | 7865 |
Today
— stars today
This week
+1 stars this week
This month
— stars this month
Momentum
1.0
growth rate 0.01%/day
<img src="./chatgpt.png" width="450px"></img>
*<a href="https://openai.com/blog/chatgpt/">official chatgpt blogpost</a>*
## PaLM + RLHF - Pytorch (wip)
Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture. Maybe I'll add retrieval functionality too, à la <a href="https://github.com/lucidrains/RETRO-pytorch">RETRO</a>
If you are interested in replicating something like ChatGPT out in the open, please consider joining <a href="https://discord.gg/xBPBXfcFHd">Laion <img alt="Join us on Discord" src="https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white"></a>
Potential successor: <a href="https://arxiv.org/abs/2305.18290">Direct Preference Optimization</a> - all the code in this repo becomes ~ binary cross entropy loss, < 5 loc. So much for Reward models and PPO
## FAQ
- Does this contain a model for inference?
There is no trained model. This is just the ship and overall map. We still need millions of dollars of compute + data to sail to the correct point in high dimensional parameter space. Even then, you need professional sailors (like Robin Rombach of Stable Diffusion fame) to actually guide the ship through turbulent times to that point.
## Community
<a href="https://carper.ai/">CarperAI</a> had been working on <a href="https://github.com/CarperAI/trlx">an RLHF framework</a> for large language models for many months prior to the release of ChatGPT.
<a href="https://www.youtube.com/watch?v=sswA4j_IUxg">Yannic Kilcher</a> is also working on an <a href="https://github.com/LAION-AI/Open-Assistant">open sourced implementation</a>
<a href="https://www.youtube.com/watch?v=SWwQ3k-DWyo">AI Coffeebreak w/ Letitia</a> | <a href="https://www.youtube.com/watch?v=NpmnWgQgcsA">Code Emporium</a> | <a href="https://www.youtube.com/watch?v=_MPJ3CyDokU">Code Emporium Part 2</a>
## Appreciation
- <a href="https://stability.ai/">Stability.ai</a> for the generous sponsorship to work on cutting edge artificial intelligence research
- <a href="https://huggingface.co/">🤗 Hugging Face</a> and <a href="https://carper.ai/">CarperAI</a> for penning the blog post <a href="https://huggingface.co/blog/rlhf">Illustrating Reinforcement Learning from Human Feedback (RLHF)</a>, and the former also for their <a href="https://huggingface.co/docs/accelerate/index">accelerate</a> library
- <a href="https://github.com/kisseternity">@kisseternity</a> and <a href="https://github.com/taynoel84">@taynoel84</a> for the code review and finding bugs
- <a href="https://github.com/conceptofmind">Enrico</a> for integrating <a href="https://arxiv.org/abs/2205.14135">Flash Attention</a> from Pytorch 2.0
- [bycloud](https://www.youtube.com/@bycloudAI) for his educational video for [Reasoning with Exploration](https://www.youtube.com/watch?v=uOrJUksvIhs)
## Install
```bash
$ pip install palm-rlhf-pytorch
```
## Usage
First train `PaLM`, like any other autoregressive transformer
```python
import torch
from palm_rlhf_pytorch import PaLM
palm = PaLM(
num_tokens = 20000,
dim = 512,
depth = 12,
flash_attn = True # https://arxiv.org/abs/2205.14135
).cuda()
seq = torch.randint(0, 20000, (1, 2048)).cuda()
loss = palm(seq, return_loss = True)
loss.backward()
# after much training, you can now generate sequences
generated = palm.generate(2048) # (1, 2048)
```
Then train your reward model, with the curated human feedback. In the original paper, they could not get reward model to be finetuned from a pretrained transformer without overfitting, but I gave the option to finetune with `LoRA` anyways, since it is still open research.
```python
import torch
from palm_rlhf_pytorch import PaLM, RewardModel
palm = PaLM(
num_tokens = 20000,
dim = 512,
depth = 12,
causal = False
)
reward_model = RewardModel(
palm,
num_binned_output = 5 # say rating from 1 to 5
).cuda()
# mock data
seq = torch.randint(0, 20000, (1, 1024)).cuda()
prompt_mask = torch.Excerpt of 13,939 characters
Read on GitHub
Phil Wang · United States
145
Enrico Shippole · Teraflop AI
10
6
Ikko Eltociear Ashimine · Japan
2
Eric Alcaide · Switzerland
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:846df491fb1a5e3e, topic:reinforcement-learning, desc:reinforcement learning, readme:reinforcement learning
matched fp:846df491fb1a5e3e, topic:deep-learning
matched fp:846df491fb1a5e3e, readme:finetune, readme:lora, name:rlhf