Replying to @Tom_A_Lynch
Remember when teenage engineering designed Raven R for Baidu
When fine-tuning small LLMs, it's interesting to see which anchors are established.
For the most part, at 4-bit quantization, very little personality is preserved, but certain functions do seem to peek through.
This fine-tuning is made from my entire ChatGPT conversation dataset of around 9k conversations.
I've made a new version of my video-generating program:
1. scrapes the internet for articles
2. Generates a script
3. Generates a series of images using Nano-banana
4. Generates the voice using Eleven Labs
5. Combines the elements and exports mov













