HomeThe differenceHow it's builtThe numbersTalk to her
lanox

newchar-leveltrained from scratch

I trained a model

No big pretrained brain and nobody else's voice mixed in.Just her messages.

Talk to her
How it's built
Lanox, an anime girl with white hair and cat ears
waking her up

01 / the difference

Same text.
Two replies.

Big models are trained to be pleasant to everyone. This one was trained on a single person, so it answers the way she does — short, dry, and entirely unbothered.

any assistant

trained on everyone

how are you

I'm doing great, thank you so much for asking! 😊 How about you? I hope your day is going wonderfully!

chars
101
emoji
1
!
2
tone
polite
L

lanox

trained on one person

how are you

chars
9
emoji
0
!
0
tone
blunt

// illustrative replies, written to show the style each one is trained toward

02 / the build


The corpus is small and the compute is free, so the training is stubborn about it: learn to be fluent, then learn to be her, then get nudged toward replies a judge can't tell apart from the real ones.

  1. 01 · step 0

    From scratch

    No pretrained weights, no borrowed vocabulary. A hand-written decoder-only transformer that learns text one character at a time — starting from noise.

    herq3l;z h.. ekkvn xo@ ,,m

  2. 02 · design bet

    One model, not two

    Style and content share the same weights. Generate-then-restyle pipelines give you a fluent stranger wearing her outfit. This doesn't do that.

    herhhe yu aa wht th.. ok ok

  3. 03 · phase 1

    Pretrain

    First it learns to be fluent at all, on synthetic casual dialogue from a teacher model. A plausible guess at how people text — explicitly not her, yet.

    herHey! I'm doing great, thanks for asking. How about you? 😊

  4. 04 · phase 2

    Fine-tune

    Then it meets the real timeline, at a much lower learning rate, so her actual habits overwrite the generic fluency instead of erasing it.

    herhey whats up

  5. 05 · phase 3

    RL with a judge

    A GRPO loop samples a group of replies, an LLM judge scores them, the policy leans toward the good ones. Only her side is trainable — the other side is always real history.

    herwhy are you asking me that

  6. 06 · the leash

    KL penalty

    A divergence penalty against the frozen Phase 2 model keeps her voice from drifting into reward-hacked mush. It's allowed to improve, not allowed to become someone else.

    heridk you tell me

The judge does not reward warmth.

// if she'd be blunt, blunt scores full marks. no bonus points for
// affection she wouldn't actually send. that's the whole trick.

the training stack

  • PyTorch

    hand-written model + loop

  • char-level

    398 symbols, emoji included

  • MiniGPT

    decoder-only transformer

  • AdamW

    cosine LR, warmup

  • GRPO

    group-relative RL

  • LLM judge

    scores realism, not warmth

  • Qwen

    fallback judge

  • KL penalty

    keeps her voice hers

03 / inside the net

The whole thing,
in numbers.

Small enough to train on a free Colab T4 — and honestly small enough to run without a GPU at all. That's the point.

her whole timeline, as raw text
Characters seen
context → her reply
Training pairs
emoji = one char
Vocab size
tiny, on purpose
~15M
Parameters

she's typing

Go on. Text her.

Fair warning — she was trained to be herself, not to be nice about it.

Open the playground