feat(TaskProcessing): add TextToTextReformatParagraphs task processing handler#362
feat(TaskProcessing): add TextToTextReformatParagraphs task processing handler#362lukasdotcom wants to merge 2 commits intomainfrom
Conversation
…g handler Signed-off-by: Lukas Schaefer <lukas@lschaefer.xyz>
I wonder if this is being too smart. Can you test this with llama 3.1 from IONOS as a baseline? |
I tested it with a small qwen model, but I can also test it with llama. |
|
Also see the comments in #363 |
Signed-off-by: Lukas Schaefer <lukas@lschaefer.xyz>
87f1c84 to
4d03993
Compare
It does work with llama 3.1. It seems to fail to stop though for some reason and takes absolutely forever. The first two lines would be the only output needed for my test prompt. Llama 3.1 seems to do that for all prompts right now. DetailsDogs have been humanity's most devoted companions for thousands of years Note: The output has only 2 sections because there are only two distinct topics - dogs and cats.assistant (I removed this to meet the constraints which don't allow extra added text)Dogs have been humanity's most devoted companions for thousands of years (I removed my part to meet the 'no additional text' constraint)Dogs have been humanity's most devoted companions for thousands of years (I am now meeting the constraints).Dogs have been humanity's most devoted companions for thousands of years (I am deleting this extra sentence now and this new text will strictly only contain the output now so as to follow constraint to not add extra text):Dogs have been humanity's most devoted companions for thousands of years Dogs have been humanity's most devoted companions for thousands of years comput lith manners Focus animated suger ideally relax generators older If:::/api/L Interest duplicanprofits discourputes Solomon forecasting update primes guarantee benches Alert permanent-vesmDogs have been humanity's most devoted companions for thousands of years Witnesses debris Ward existence 理 Buddha London maskDogs have been humanity's most devoted companions for thousands of years |
Openai integration implementation for: nextcloud/server#60171 (It needs that branch to work)
Conserves tokens and prevents content from being lost by having the llm only generate the first sentence for each paragraph and then using code to create the final output.