OK : Next Few Models I will Perform Traiing With NO PROMPTS ! ... This is only to produce the model for LEADERBOARD : ( NOT FOR USAGE )
As you know the leader board uses MULTIPLE CHOICE @
I DID NOT TRAIN FOR MULTIPLE CHOICE ! as why would an AI need to do Multiple choice exam ! , i call this out of domain knowledge ! as well as out of domain usages !
SO:
If we need a model which is dead ! we can remove the prompt in training and concentrate on multiple choice only ! ( it does not matter the questions as the model just needs to learn the GAME! )
have found that on new tasks it takes a warm up to align itself to the expected output formats even in training this can take exen 1000 samples in strange cases and other only 50-100! Since this is a new task style it will take some time ( ie 1000 )
Mu8ltiple choice is actually a game style QA , so the more examples of tricky and misleading questions and selelcting the correct , or Idea that the questioner was portraying or expect , then we can proceed to learn ! so IT doe not give the model intleligence ! , and this is because the answers are closed ! , and the answer is often frame as A, B,C,D but a good model will give the answer and reason !, as well as repeat the whole answer lol providing that enriched formatting expected in natural language interfaces ! so essentially you have to DUMB DOWN the model to accept these questions and provide one Character Answers !@ ( going against the model , imagine in that case you might have to say the max tokenn out is 1)
what a croc !
hence great model did not imporve at all despite trainign verbatum in the leaderboard datasets of which they do ! also !!
( bad pop knowledge to downgrade your models ) -- - -- -- We shall just remove and see what happens ( it will be on the chat ml as it os the format best used to force a model just to output yes/no answers !)
- Developed by: LeroyDyer
- License: apache-2.0
- Finetuned from model : LeroyDyer/_Spydaz_Web_AI_ChatQA_002r1_4_BIT
This mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.
- Downloads last month
- 17
Model tree for LeroyDyer/_Spydaz_Web_AI_ChatQA_003
Base model
LeroyDyer/_Spydaz_Web_AI_ChatQA_002_4_BIT