Fine Tuning not responding with tuned data

API
,
1

Intention of creating bizarre and unique 3 prompts is to test and confirm the model is tuned correctly.

Can you please suggest why chatGPT CLI is not responding with the completions I tuned it with?

Prepared JSONL file
{“prompt”:“what is abakaas?”,“completion”:" ABAKAAS is a platform star that drives RABACSA on a star’s mooon for 243 days.“}
{“prompt”:“what is bdbasdbasd?”,“completion”:” BDBASDBASD is the formula used to capture comet’s tail praschts in real time.“}
{“prompt”:“what is dsdfsdfe?”,“completion”:” DSDFSDFE is the theory behind writing Torums in spanish in 1930s."}

CLI Create Model
$ openai api fine_tunes.follow -i ft-< ID >

[2023-05-30 16:03:45] Created fine-tune: ft-< ID >

[2023-05-30 16:05:12] Fine-tune costs $0.01

[2023-05-30 16:05:13] Fine-tune enqueued. Queue number: 0

[2023-05-30 16:15:15] Fine-tune started

[2023-05-30 16:18:30] Completed epoch 1/4

[2023-05-30 16:18:31] Completed epoch 2/4

[2023-05-30 16:18:32] Completed epoch 3/4

[2023-05-30 16:18:33] Completed epoch 4/4

[2023-05-30 16:19:11] Uploaded model: davinci:ft-personal:test02-2023-05-30-23-19-11

[2023-05-30 16:19:12] Uploaded result file: file-<Result_ID >

[2023-05-30 16:19:12] Fine-tune succeeded

Job complete! Status: succeeded ![:tada:]

Try out your fine-tuned model:

openai api completions.create -m davinci:ft-personal:test02-2023-05-30-23-19-11 -p <YOUR_PROMPT>

CLI Test

$ openai api completions.create -m davinci:ft-personal:test02-2023-05-30-23-19-11 -p “what is dsdfsdfe?”

what is dsdfsdfe?!!?!!!’ was the question on everyone’s mind, myself included certain characters

$ openai api completions.create -m davinci:ft-personal:test02-2023-05-30-23-19-11 -p “what is bdbasdbasd?”

what is bdbasdbasd?”, “what’s the point of using bdbas

$ openai api completions.create -m davinci:ft-personal:test02-2023-05-30-23-19-11 -p “what is abakaas?”

what is abakaas? how does it work?

ABAKAAS is a platform for funding

2

You are missing the stop sequence:

  • Use a separator at the end of the prompt, e.g. \n\n###\n\n. Remember to also append this separator when you eventually make requests to your model.

Also, look into embeddings for this kind of stuff.

3

@curt.kennedy Based on your recommendation I created a new model with the following changes, but it did not resolve the issue .

Prepared JSONL file
{“prompt”:“what is abakaas?\n\n###\n\n”,“completion”:" ABAKAAS is a platform star that drives RABACSA on a star’s mooon for 243 days.\n"}
{“prompt”:“what is bdbasdbasd?\n\n###\n\n”,“completion”:" BDBASDBASD is the formula used to capture comet’s tail praschts in real time.\n"}
{“prompt”:“what is dsdfsdfe?\n\n###\n\n”,“completion”:" DSDFSDFE is the theory behind writing Torums in spanish in 1930s.\n"}

CLI Create Model
$ openai api fine_tunes.follow -i ft-< ID >
[2023-05-30 22:02:21] Created fine-tune: ft-< ID >
[2023-05-30 22:03:40] Fine-tune costs $0.01
[2023-05-30 22:03:41] Fine-tune enqueued. Queue number: 0
[2023-05-30 22:03:44] Fine-tune started
[2023-05-30 22:06:54] Completed epoch 1/4
[2023-05-30 22:06:55] Completed epoch 2/4
[2023-05-30 22:06:56] Completed epoch 3/4
[2023-05-30 22:06:57] Completed epoch 4/4
[2023-05-30 22:07:40] Uploaded model: davinci:ft-personal:test03-2023-05-31-05-07-39
[2023-05-30 22:07:41] Uploaded result file: file-< Result ID >
[2023-05-30 22:07:41] Fine-tune succeeded

Job complete! Status: succeeded :tada:
Try out your fine-tuned model:

openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p <YOUR_PROMPT>
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “what is abakaas?\n\n###\n\n”
what is abakaas?\n\n###\n\n\tlão se’ihabakaas nyu waawã!:
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “what is bdbasdbasd?\n\n###\n\n”
wwhat is bdbasdbasd?\n\n###\n\n’) def 3_S1(): import os print '### what\ssss
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “wwhat is bdbasdbasd?”
wwhat is bdbasdbasd?robowt essa sueme daserggket jow dbhlza
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “wwhat is obamacare?”
wwhat is obamacare? A: Obamacare is brilliant. And it has set itself up for sweet continuous tax
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “what is dsdfsdfe?\n\n###\n\n”
what is dsdfsdfe?\n\n###\n\n"$dsdfsdfe #~ “dsdfdsfd\n
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “what is dsdfsdfe?\n”
what is dsdfsdfe?\n”);

44

45 double num = 0;

46
$ openai api completions.create -m davinci:ft-personal:test03-2023-05-31-05-07-39 -p “what is bdbasdbasd?\n\n###\n\n”
what is bdbasdbasd?\n\n###\n\n" dnet += \ “,hdwrdg\n” dnet +=

1 Like
4

Sorry, I am not answering this post, but I want to know how this post would be addressed or resolved.

I really don’t understand why having “\n\n###\n\n” and not having “\n\n###\n\n” would yield different results. This indirectly tells me that “H ello, \n how are you?” and “Hello How Are You” are considered different. (We know “\n” is a newline character. However, we also instantly understand that they have the same meaning.)

Could the AI be intelligent enough to recognize irrelevant data, such as “\n\n###\n\n,” which people generally do not consider when reading a text? Has the fine-tuning technique not advanced sufficiently to address this issue?

5

you can use any stop sequnence like ### , **** anything
but make sure the pattern u are using for the stop sequence
should not come anywhere in middle of question and anwer

6

@vishals Like I said above, you really need to use embeddings for this stuff. Fine-tuning on 4 epochs doesn’t add knowledge to the model(s).

Here, I created a session, with simulated embeddings in the background that retrieved your content and fed it to the LLM. It answered all questions correctly.

Screenshot 2023-05-31 at 7.38.20 AM
Screenshot 2023-05-31 at 7.38.20 AM3546×1392 381 KB

1 Like