2023-04-24 Meeting 8

Fix datasets

1. Reditt validation dataset (tikai tie paraugi, kur Pauls & Laura sakrīt)

2. Prompt dataset izmantot to kuram visvairāk labeller consensus pieejams

3. Nodalīt datasets:

   1. Prompt dataset

   2. Reditt Zero-shot dataset

   3. Validaiton datasets (arī Reditt validācijai)

Dokumentēt procesu

Pierakstīt un izskaidrot - piemerus, kur atgriež ne tā kā tu prasi

Prompts

1. Binary tree metode

• positive - negative

• high / low intensity (low == neutral)

2. Tranlation metode

• Angļu valodā

• Classify

   

3. Template metodes

• This sentence is written in [token] emotional tone. Replace [token] value. ^ brīvā formā varbūt var atklāt kādu sakarību

• This sentence is written in [token] emotional tone. Replace [token] value with “positive”, “negative”, “neutral”.

4. Roleplay

• Pretend to be very emotional actor, how would you classify this sentence “….” , reply with “positive”, “negative”, “neutral”.

5. Context piemēri ,

   Pirms target daļas iekopē random 5 x 3 piemērus no katras klases - tas ir actually kā GPT strādā bez HFRL

6. Probablity pievienot reply

   Positive: 77%