{"slug": "using-jev-as-a-teacher-to-help-an-slm-write-better-stories", "title": "Using Jev as a teacher to help an SLM write better stories", "summary": "A developer trained a roughly 14-million-parameter small language model on the TinyStories dataset and paired it with Jev, a model that judges text by answering yes/no questions with calibrated confidence, to improve its story generation. The tiny writer drafts one sentence at a time and rewrites any sentence Jev flags as repetitive, off-topic, or unmotivated, correcting an example line about \"a band of different colors\" into \"she had a big, red ball that she loved.\" The experiment tests how small a model can be and still produce coherent stories, given that large chatbot models carry tens or hundreds of billions of parameters.", "body_md": "# Giving a tiny story-writer a teacher\n\nFor a while now I've been playing with a slightly odd question: **how small can an\nAI be and still tell a story?**\n\nThe AI systems everyone talks about — the ones behind chatbots — are enormous.\nThey have tens or hundreds of *billions* of internal settings (people call these\n\"parameters\" — think of them as tiny knobs the model tunes as it learns). That\nsize is why they're expensive to run and impossible to fit on a phone or a toy.\nI wanted to go the other way. How much can you strip away and still get something\nthat writes a real, if wobbly, little story?\n\n## A storyteller the size of a rounding error\n\nSo I built one. I shrank the vocabulary down to the simple words a small child\nuses, cut the model's size again and again, and trained it on **TinyStories** — a\ncollection of very simple tales written with a toddler-sized vocabulary, made\nexactly for this kind of experiment. The result is a storyteller with about **14\nmillion** parameters. That sounds like a lot until you remember the big models\nhave *a hundred thousand times* more.\n\nAnd it works — sort of. It writes recognizable little stories about girls and boys and balls and parks. They're charming. They're also, frankly, not very good. It loses the thread, repeats itself, introduces a character and forgets them, has things happen for no reason. Here's a real example of what it produced on its own:\n\n…they also liked to pretend they were explorers and pretend they were explorers. …ben had a sad face and a hole in its fur…\n\nA stutter (\"pretend they were explorers\" twice), and a friend who somehow has fur. Fun to read, but not a story you'd want read to you.\n\nThat's roughly where you'd expect a model this tiny to top out. But I didn't want to stop there.\n\n## The insight: a child doesn't write alone\n\nHere's the thing that nagged at me. A four-year-old tells stories with a *tinier*\nvocabulary than my model — fewer than two thousand words — and their stories are\noften better. How?\n\nPart of the answer is that **a child rarely writes alone.** There's usually a\nparent or a teacher nearby, gently nudging: *\"Wait — why did the horse start\nsinging?\"* … *\"What did the forest look like?\"* … *\"You said the leaves were soft\n*and* sharp — which is it?\"* The child doesn't need those questions written into\ntheir head. They just need someone to ask them at the right moment, and they\ncorrect themselves.\n\nSo I wondered: **what if my tiny writer had a teacher too?**\n\n## Enter the teacher\n\nThe teacher I reached for is a model called **Jev**. Jev is unusual: it doesn't\nwrite anything. What it does is *judge*. You show it a piece of text and ask it\nprecise questions — \"Does this sentence follow from what came before?\", \"Is this\ntoo repetitive?\", \"Did this event happen for a reason the story gives?\" — and it\nanswers each one with a **calibrated confidence**, a number saying how sure it is.\nIt's less like a writer and more like a very fast, very consistent reading teacher\nwho only ever answers yes/no questions, honestly.\n\nThat turns out to be exactly the shape of a teacher you can put in a loop. So here\nis what my tiny writer now does. It writes **one sentence**. Jev reads the whole\nstory so far and checks it against a handful of those questions. If everything's\nfine, the writer moves on. If Jev flags something — the sentence drifts off topic,\nor repeats a word, or an event came out of nowhere — the writer simply **tries that\nsentence again**, and again, until it passes. Then it writes the next one, with the\nteacher watching the whole time.\n\n## It actually works\n\nHere's a single correction, caught live. The writer produced:\n\n✗ *she wanted a band of different colors.*\n\nJev flagged it — not for grammar (it's fine) and not for being off-topic, but for\nbeing **unmotivated**: a vague wish that leads nowhere. (\"A band of different\ncolors\" — you can't quite picture it, and nothing follows from it.) So the writer\ntried again:\n\n✓ *she had a big, red ball that she loved to play with.*\n\nA concrete thing, one the rest of the story can actually use. The teacher's own coherence rating for the story jumped from a 2 to a 3 out of 3.\n\nMultiply that across every sentence and the difference is night and day. Remember the stuttering, fur-covered mess from before? With the teacher in the loop, the same little model wrote this:\n\nonce 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.\n\nThat's a *real little arc* — a setup, a mishap, a reaction — instead of a stutter\nand a friend with fur. Same tiny brain. The only thing that changed is that\nsomeone was asking it the right questions as it went.\n\n## Why I think this is a big deal\n\nSit with what just happened: a model a hundred-thousandth the size of a frontier\nsystem produced a genuinely better story — not by getting bigger, but by getting\n**coached.** The intelligence didn't all have to live *inside* the writer. Some of\nit lived in the teacher standing next to it.\n\nThat opens up two directions I find genuinely exciting.\n\nThe first is about **doing more with less.** If a small, cheap model plus a good\nteacher can punch this far above its weight, you don't always need the giant\nexpensive model. You need a decent little one and a sharp critic — and a critic\nthat only answers focused questions can be small and cheap too. (I want to be\nhonest about the ceiling: the teacher can only ever pick from sentences the little\nwriter is *capable* of producing — it can't put words in its mouth it could never\nhave found. It makes the writer reliably do its best, not become something else.\nBut \"reliably its best\" is a large improvement.)\n\nThe second is about **people, not models.** Look again at that loop — a learner\nproduces something, a patient teacher reads it and asks exactly the right question\nat exactly the right moment, and the learner revises. That's not just how you coach\na tiny AI. **That's what good teaching is.** The same idea — a tireless, consistent\ntutor that reads a student's own writing and nudges with the right question instead\nof just marking it wrong — could sit inside education products and help real\nchildren write better. That's the version of this I most want to explore.\n\n## What's next: learning, not just redoing\n\nThere's one honest limitation in what I've described. Right now, when the teacher\nflags a sentence, the writer just **rewrites it** — in that moment. It doesn't\n*remember* the lesson. Ask it to write a fresh story tomorrow with no teacher, and\nit'll make the same kinds of mistakes again.\n\nReal learning is the opposite: you internalize the feedback so you stop needing it.\nI've taken a first step in that direction — I fed the teacher-corrected stories\n*back into* the little model as practice, and its solo writing did measurably\nimprove (it even learned to actually *finish* its stories, which it rarely did\nbefore). But that's just the beginning.\n\nThe frontier I'm heading toward is a proper **feedback curriculum**: not a teacher\nthat corrects the same mistake forever, but one that turns each correction into a\nlesson the writer keeps — the way a child eventually stops needing to be asked \"why\ndid that happen?\" because they've learned to ask it themselves. Right now my little\nwriter has a teacher. Next, I want to give it an education.\n\nMore soon.", "url": "https://wpnews.pro/news/using-jev-as-a-teacher-to-help-an-slm-write-better-stories", "canonical_source": "https://blog.trulm.com/posts/tiny-story-writer-with-a-teacher/", "published_at": "2026-09-20 17:02:00+00:00", "updated_at": "2026-09-20 17:22:59.276055+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-tools", "natural-language-processing"], "entities": ["Jev", "TinyStories"], "alternates": {"html": "https://wpnews.pro/news/using-jev-as-a-teacher-to-help-an-slm-write-better-stories", "markdown": "https://wpnews.pro/news/using-jev-as-a-teacher-to-help-an-slm-write-better-stories.md", "text": "https://wpnews.pro/news/using-jev-as-a-teacher-to-help-an-slm-write-better-stories.txt", "jsonld": "https://wpnews.pro/news/using-jev-as-a-teacher-to-help-an-slm-write-better-stories.jsonld"}}