File size: 1,568 Bytes
a426e5b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ae7ecb2
a426e5b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
datasets:
- HuggingFaceH4/no_robots
language:
- en
license: cc-by-nc-4.0
---

# Good Robot 2 🤖

The model "Good Robot" had one simple goal in mind: to be a good instruction-following model that doesn't talk like ChatGPT.

Built upon the Mistral 7b 0.2 base, this model aims to provide responses that are as human-like as possible, thanks to some DPO training using the (for now, private) `minerva-ai/yes-robots-dpo` dataset.


HuggingFaceH4/no-robots was used as the base for generating a custom dataset to create DPO pairs.

It should follow instructions and be generally as smart as a typical Mistral model - just not as soulless and full of GPT slop.

Changes from the original [good-robot](https://huggingface.co/kubernetes-bad/good-robot) model:
  - Mistral 7b-0.2 base (32k native context, no SWA)
  - ChatML prompt format
  - Trained using GaLore method

## Prompt Format:

ChatML
```
<|im_start|>system
System message
<|im_start|>user
User message<|im_end|>
<|im_start|>assistant
```

## Credits:
Model made in collaboration with [Gryphe](https://huggingface.co/Gryphe).

## Training Data:
- [HuggingFaceH4/no_robots](https://huggingface.co/datasets/HuggingFaceH4/no_robots)
- [MinervaAI/yes-robots-dpo](https://huggingface.co/MinervaAI)
- private datasets with common GPTisms


## Limitations:

While I did my best to minimize GPTisms, no model is perfect, and there may still be instances where the generated content has GPT's common phrases - I have a suspicion that's due to them being engrained into Mistral model itself.

## License:
cc-by-nc-4.0