Testing ChatGPT

 So I decided to test ChatGPT to see if it lived up to the hype. Yes it does.

However it seems limited to about 8000 chars input. So you can't give it wikipedia and ask for a summary.

But it's really impressive.

I asked it to summarise 4 pages of an article on conspiracy theories. Then I addressed it in Afrikaans and German. No problems. Even random "politeness" things we do in speaking to a human, it gives appropriate responses to, even though they do not contain significant propositional content. I am impressed.

See screenshots below.






Popular posts from this blog

Hallucinations/inaccuracies

 You have to test LLMs individually for hallucinations and inaccuracies. These below are attempts at summarising my first academic article. Useless! john@john-GF63-Thin-11SC:~/ai$ ./scripts/ollama_summarise_one.sh SAMPLES/hawaiiarticle.txt  Sure, here's the paragraph you requested: >The problem with some of the analyses of Libet is that they make it look  like the details were complicated. However, I do believe I agree with  McCrone that Libet's data is sloppier than people often realize. Glynn  made that point a long time ago and it was the main point of my plots in  the Tucson III Proceedings. john@john-GF63-Thin-11SC:~/ai$ ./scripts/ollama_summarise_one.sh SAMPLES/hawaiiarticle.txt  Sure, here is the paragraph you requested: >Libet showed that a thalamic stimulus requires a duration of about 250  msec to be felt whereas a 20 msec skin stimulus is adequate. This shows  that the intensity of the stimulus plays an important role in dete...
Read more

Access to model mistralai/Mistral-Nemo-Base-2407 is restricted. You must be authenticated to access it.

 If you get an error like Access to model mistralai/Mistral-Nemo-Base-2407 is restricted. You must be authenticated to access it. Then pip install huggingface_hub huggingface-cli login it will ask you for a token go to https://huggingface.co/settings/tokens it will have an option to generate a token make sure you click the checkboxes for APIs generate the token copy it go back to the terminal at the top paste.
Read more

Building my own LLM

Image
Yesterday I started building my own LLM because I want to avoid licensing issues for basic chatbot usages. It's super slow training these things. I won't post code here because I plan to commercialise it, but here's the progress screenshot: I've set it to use the following parameters. Sadly my GPU is rather small so it is taking ages. # Training hyperparameters TEMPERATURE=0.7 TOP_K=40 TOP_P=0.95 REPEAT_PENALTY=1.1 MAX_NEW_TOKENS=200 BLOCK_SIZE=32 BATCH_SIZE=8 NUM_EPOCHS=5 LEARNING_RATE=0.001 CHECKPOINT_INTERVAL=1 USE_GPU=true # File and directory settings MODEL_DIR=models/ CORPUS_DIR=repo/ LOG_DIR=logs/ BACKUP_DIR=bak/ TOKENIZER_PATH= # Internal settings EVAL_INTERVAL=100 EVAL_ITERS=100 SEED=1337 INPUT_FILE=input.txt
Read more