syncopate-machine ⚡

Tiny Burn transformer for Dancing With My Code chat and his game. It predicts action IDs, then the website response graph turns those actions into short Thai/English messages and mood GIFs. Small brain, fast vibes.

user text -> website classifier -> [action_id, lang_id, SEP]
          -> syncopate-machine -> next action IDs
          -> response graph -> message + mood GIF

Quick Start 🚀

Build the action data:

python examples\build_action_data.py --offline

Train on CUDA:

cargo run --release --features cuda --example train_action_model -- `
  --steps 20000 --batch-size 64 --lr 0.003 `
  --checkpoint-dir runs/action-v4-d16-clean

Ship these into the website:

dancing-with-my-code-v2/assets/model-personal-v2.mpk
dancing-with-my-code-v2/assets/model-config-personal-v2.json

Action Vocabulary 📝

15 tokens — control codes, action intents, and language tags:

ID	Token	Role
0	PAD	padding
1	SOS	start of sequence
2	EOS	end of sequence
3	SEP	separator
4	Unknown	unclassified input
5	Greeting	สวัสดี / hello vibes
6	Farewell	บ๊ายบาย / goodbye
7	Insult	ด่า / roast mode
8	Programming	code talk
9	Identity	who am I?
10	Resume	แนะนำตัว / intro
11	Links	link drops
12	Course	course info
13	TH	Thai language tag
14	EN	English language tag

Model Vibe 🧠

Setting	Value
Task	causal action-sequence prediction
Vocab	15 action/lang/control IDs
Sequence length	16
Layers	1
d_model	16
Attention	causal scaled dot-product softmax
Attention heads	4
KV heads	1
FFN	SwiGLU, width 16
Position encoding	RoPE
Norm	RMSNorm
Output projection	tied to embedding
Params	1,696

Current Checkpoint 💾

The website uses:

runs/action-v4-d16-clean/final.mpk

Current validation metrics:

Metric	Value
Validation loss	0.6742
Validation perplexity	1.96
Validation accuracy	71.15%

Training Loss 📉

Raw loss is noisy because the batches are tiny. The pink line is an 80-step smooth, which shows the real shape: hard drop early, then slow grind.

Browser Runtime 🌐

Mode	Behavior
auto	try WebGPU, fall back to CPU
gpu	require WebGPU
cpu	force CPU/Flex backend

MIT. Break it, fix it, ship it. ✨