### Guanaco SuperCOT
Guanaco SuperCOT is trained with the aim of making LLaMa follow prompts for Langchain better, by infusing chain-of-thought datasets, code explanations and instructions, snippets, logical deductions and Alpaca GPT-4 prompts. It's also an advanced instruction-following language model built on Meta's LLaMA 33B model. Expanding upon the initial 52K dataset from the Alpaca model, an additional 534,530 entries have been incorporated, covering English, Simplified Chinese, Traditional Chinese (Taiwan), Traditional Chinese (Hong Kong), Japanese, Deutsch, and various linguistic and grammatical tasks. This wealth of data enables Guanaco to perform exceptionally well in multilingual environments.
It uses a mixture of the following datasets:
[https://huggingface.co/datasets/QingyiSi/Alpaca-CoT](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT)
- Chain of thought QED
- Chain of thought Aqua
- CodeAlpaca
[https://huggingface.co/datasets/neulab/conala](https://huggingface.co/datasets/neulab/conala)
- Code snippets
[https://huggingface.co/datasets/yahma/alpaca-cleaned](https://huggingface.co/datasets/yahma/alpaca-cleaned)
- Alpaca GPT4
- [https://huggingface.co/datasets/JosephusCheung/GuanacoDataset](https://huggingface.co/datasets/JosephusCheung/GuanacoDataset)
- Guacano
- [https://huggingface.co/timdettmers/guanaco-33b](https://huggingface.co/timdettmers/guanaco-33b)
- Guacano 33b LoRa
1\. Prompting
-------------------------
You should prompt the LoRA the same way you would prompt Alpaca or Alpacino.
The new format is designed to be similar to ChatGPT, allowing for better integration with the Alpaca format and enhancing the overall user experience.
Instruction is utilized as a few-shot context to support diverse inputs and responses, making it easier for the model to understand and provide accurate responses to user queries.
The format is as follows:
```
### Instruction:
User: History User Input
Assistant: History Assistant Answer
### Input:
System: Knowledge
User: New User Input
### Response:
New Assistant Answer
```
This structured format allows for easier tracking of the conversation history and maintaining context throughout a multi-turn dialogue.
```
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
### Input:
### Response:
```
Remember that with lower parameter sizes, the structure of the prompt becomes more important. The same prompt worded differently can give wildly different answers. Consider using the following suggestion suffixes to improve output quality:
- "Think through this step by step"
- "Let's think about this logically"
- "Explain your reasoning"
- "Provide details to support your answer"
- "Compare and contrast your answer with alternatives"
2\. Role-playing support:
-------------------------
Guanaco now offers advanced role-playing support, similar to Character.AI, in English, Simplified Chinese, Traditional Chinese, Japanese, and Deutsch, making it more versatile for users from different linguistic backgrounds.
Users can instruct the model to assume specific roles, historical figures, or fictional characters, as well as personalities based on their input. This allows for more engaging and immersive conversations.
The model can use various sources of information to provide knowledge and context for the character's background and behavior, such as encyclopedic entries, first-person narrations, or a list of personality traits.
The model will consistently output responses in the format "Character Name: Reply" to maintain the chosen role throughout the conversation, enhancing the user's experience.
3\. Rejection of answers and avoidance of erroneous responses:
--------------------------------------------------------------
The model has been updated to handle situations where it lacks sufficient knowledge or is unable to provide a valid response more effectively.
Reserved keywords have been introduced to indicate different scenarios and provide clearer communication with the user:
* NO IDEA: Indicates that the model lacks the necessary knowledge to provide an accurate answer, and will explain this to the user, encouraging them to seek alternative sources.
* FORBIDDEN: Indicates that the model refuses to answer due to specific reasons (e.g., legal, ethical, or safety concerns), which will be inferred based on the context of the query.
* SFW: Indicates that the model refuses to answer a question because it has been filtered for NSFW content, ensuring a safer and more appropriate user experience.
4\. Continuation of responses for ongoing topics:
-------------------------------------------------
The Guanaco model can now continue answering questions or discussing topics upon the user's request, making it more adaptable and better suited for extended conversations.
The contextual structure consisting of System, Assistant, and User roles allows the model to engage in multi-turn dialogues, maintain context-aware conversations, and provide more coherent responses.
The model can now accommodate role specification and character settings, providing a more immersive and tailored conversational experience based on the user's preferences.
It is important to remember that Guanaco is a 7B-parameter model, and any knowledge-based content should be considered potentially inaccurate. We strongly recommend providing verifiable sources, such as Wikipedia, for knowledge-based answers. In the absence of sources, it is crucial to inform users of this limitation to prevent the dissemination of false information and to maintain transparency.
### Citations
Alpaca COT datasets
```
@misc{alpaca-cot,
author = {Qingyi Si, Zheng Lin },
school = {Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China},
title = {Alpaca-CoT: An Instruction Fine-Tuning Platform with Instruction Data Collection and Unified Large Language Models Interface},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/PhoebusSi/alpaca-CoT}},
}
```
Stanford Alpaca
```
@misc{alpaca,
author = {Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto },
title = {Stanford Alpaca: An Instruction-following LLaMA model},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/tatsu-lab/stanford_alpaca}},
}
```
Google FLAN
```
@inproceedings{weifinetuned,
title={Finetuned Language Models are Zero-Shot Learners},
author={Wei, Jason and Bosma, Maarten and Zhao, Vincent and Guu, Kelvin and Yu, Adams Wei and Lester, Brian and Du, Nan and Dai, Andrew M and Le, Quoc V},
booktitle={International Conference on Learning Representations}
}
```