Using Jev as a teacher to help a 14M model write stories
Jev acts as a zero-shot classifier that scores stories from a TinyStories-based small model and suggests revisions; after rewriting, coherence rose from 2.0 to 3.0 and stories gained a complete beginning, middle and end.
The biggest advantage of Jev is that its a zero shot classifier. Biggest advantage is that we can quickly try experiments. So I tried a fun one: Using Jev as a teacher for a story writer. I have been experimenting with creating a story writer LLM, reducing vocabulary, reducing parameter size etc. I have been able to build a 14M parameter story teller based on Tiny stories dataset. As expected, the stories are not great, but good fun to try. But I wanted to improve on this. The insight I was working on was that a kid uses just 2000 words(lesser than that) to tell stories. Now, what if when she writes the stories, there is a teacher guiding the story writing, can she correct herself. And what if that teacher was Jev. With jev we can give those questions and Jev gives suggestions based on the story so far, then we ask the student to rewrite. Turns out it works. ✗ (motivated 0.32): she wanted a band of different colors. ✓ she had a big, red ball that she loved to play with. — coherence 2.0 → 3.0 Whole stories went from this: …they also liked to pretend they were explorers and pretend they were explorers. …ben had a sad face and a hole in its fur… to this: once there was a little girl named lily. she had a big, red ball that she loved to play with. one day, she went to the park to play with her ball. she was very happy and ran around all day. but then, she tripped on a rock and hurt her knee. "oh no, lily!" she cried. A real little arc — setup, mishap, reaction — instead of a stutter and a friend that has fur. The feedback from Jev and rewriting improves the stories a lot. This actually is big thing. This means we can use a small LM and then use Jev as a teacher to make it act like an LLM. We can also use Jev as a teacher in edtech products guiding students. Will explore a lot more on this. Create a instruction set so that the student can learn and write better based on feedback. Right now, it just rewrites. Will write a blog post and share soon.