alpaca.7B delivering unsatisfiying results compared to Alpaca-LoRa

[VanHallein]

Hi there,

I am getting very unsatisfying results from the alpaca.7B model compared to the Alpaca-LoRa model. I am giving alpaca.7B the following prompt, but get nothing useful out of it. However, Alpaca-LoRa handles this like a pro, similar to GPT3.5. Do you guys have any clues?

Prompt:

Instructions: Service for finding the birthday dates of all the people mentioned in the input. The output is a list of all people showing corresponding data in the following format: |John|Doe|1983/07/05|New York|USA| |Jane|Doe|1986/05/17|Paris|France|

Input: Napoleon Bonaparte[a] (born Napoleone Buonaparte; 15 August 1769 - 5 May 1821), later known by his regnal name Napoleon I, was a French military commander and political leader who rose to prominence during the French Revolution and led successful campaigns during the Revolutionary Wars. Caesar Augustus (born Gaius Octavius; 23 September 63 BC - 19 August AD 14), also known as Octavian, was the first Roman emperor; he reigned from 27 BC until his death in AD 14. Alexander III of Macedon (Ancient Greek: romanized: Alexandros; 20/21 July 356 10/11 June 323 BC), commonly known as Alexander the Great,[a] was a king of the ancient Greek kingdom of Macedon. Richard I (8 September 1157 - April 1199) was King of England from 1189 until his death in 1199.

Output:

[Vadim-Karpenko]

Sometimes it happens to me as well, it just gives an empty response. Not sure why it happens, but I managed to fix it with "An empty response is not allowed." in the instruction. By the way, could you please give me a link where I can download/try the Alpaca-LoRa model? Would appreciate it a lot.

[VanHallein]

Here are the ones I played around with. They are working fine, at least in the quality range of GPT3. The state of my alpaca.7B here is more like rambling GPT 2.0 quality.

Hugging Face Alpaca-LoRa Playground

[VanHallein]

My Problem is very much related to this #268 issue. My alpaca.7B hallucinates exactly the same way as described here.

[berksara145]

Hello, I downloaded alpaca 7B and it works fine but it is worse than I taught and my PC really struggles while trying to output. How can I download the alpaca lora model and test it in dalai? thank you

[ADANCE-STACK]

Buenas chicos tengo problemas, en la instalación baje los archivos, de LLaMA-65B-HF (boboto de hugging face), pero me parece que faltan archivos para ejecutar la instalación , alguno me puede orientar como instalar. o algun repositorio completo.

[64jcl]

For alpaca, shouldn't you use this template for this kind of query? At least for me alpaca is giving good output when I use prompts that has these ### markers:

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:
>PROMPT

### Input:
>PROMPT

### Response:

[ADANCE-STACK]

ok. gracias. entiendo que no es el sector para preguntar. y por otro lado en base a las instrcucciones no me queda claro, estoy probando una version y acorde a los prompt que tu me indicas . la manera de entrenarlo, es cuando me pone instruccion, debo ir guiandolo para que aprende de manera autonoma. o sea cuando me indica las instrucciones yo debo ir poniendole y en ese sentido el va incorporando la forma de responder seria algo asi.