Posts

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

Update on what ChatGPT is bad at doing.

OMG ChatGPT o1 is terrible. And 4.5 is not great. I keep going back to 4.o. At least it knows how to code. I've experimented now for about a year with giving ChatGPT tasks. The below criticisms apply to ChatGPT 4o . I recently got access to ChatGPT 4.5 which is better at coding and following instructions (it at least takes the instruction to give steps one at a time, seriously, whereas ChatGPT 4o almost always ignores that). Anyway, this is what 4o is bad at: Making, interpreting, processing, editing spreadsheets. You have to convert the spreadsheet to CSV first (you lose all formulas and charts). It is ok at interpreting small data sets, e.g. 20-30 bits of simple information e.g. amount of money spent per month, with each month listed. It's no good at, for example, creating a Nett Present Value spreadsheet/analysis. Editing long documents or just searching documents for specific errors. Even basic requests like checking for double spaces do not work. In one case it saw a p...
Read more

using coqui.ai and python to convert a conversation into (spoken) speech using your voice as one conversant

Install coqui (see other post here:  https://johnsaiblog.blogspot.com/2025/03/coqui-going-down-how-to-install-before.html ) Now if you save your discussion into a text file, called full_conversation . txt , in this order: speaker 1 blah blah blah speaker 2 blah blah etc etc (ie one speaker per line, no multi-line sentences), you can split the file up into sections of speech to convert to speech. Make sure each sentence is approximately 250 letters or less. Shorter is better. This script creates a folder called conversation_parts / which contains the conversation split up by speaker and by sentence.  import os import re # Input and output folder input_file = "full_conversation.txt"   # The full transcript output_folder = "conversation_parts" max_length = 250   # Maximum characters per chunk # Ensure output folder exists os.makedirs(output_folder, exist_ok=True) # Function to split text into manageable chunks def split_text(text, max_length=250):     sentence...
Read more

coqui going down - how to install before it does

 Sadly, coqui.ai is going down. It's a Text to Speech tool using python and other things. It breaks and downloads tons of dependencies if you try just installing it with pip, it's better to use conda. There is a docker option but that wants to download like 7gb of garbage.   Check that dependencies like espeak-ng and ffmpeg are installed. Installer script 1, probably won't work #!/bin/bash # Exit on error set -e echo "   Starting Coqui TTS installation..." # Ensure Miniconda is in PATH export PATH="$HOME/miniconda3/bin:$PATH" # Create and activate the Conda environment echo "   Creating Conda environment 'coqui' with Python 3.10..." conda create -n coqui python=3.10 -y source activate coqui || conda activate coqui # Install Coqui-TTS and dependencies echo "   Installing Coqui-TTS..." pip install --upgrade pip pip install coqui-tts # Install espeak-ng for phoneme-based synthesis (needed for some models) echo "   Installi...
Read more