README.md

metadata

title: WVY Tiny Liquid LM
emoji: 🌊
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 6.26.0
python_version: '3.12'
app_file: app.py
suggested_hardware: cpu-basic
short_description: Chat with a 3.31M-parameter custom liquid-style LM.

WVY Tiny Liquid LM

A Hugging Face Space for the trained TinyLiquidCausalLanguageModel checkpoint in this repository.

Model

• Parameters: 3,314,880
• Vocabulary: 8,000 BPE tokens
• Width: 192
• Blocks: 4
• Feed-forward hidden width: 512
• Causal depthwise convolution kernel: 5
• Context used by the chat app: 256 tokens
• Architecture: recurrent liquid-style state mixer + SwiGLU feed-forward layers
• Final fine-tuning loss: 0.1344
• Final fine-tuning perplexity: 1.1438

The Space loads the custom PyTorch architecture directly from model.py, restores tiny_liquid_causal_lm.pt, and uses the bundled tokenizer.json.

Files

• app.py — Gradio chat UI and autoregressive generation
• model.py — exact custom model architecture used for the checkpoint
• tiny_liquid_causal_lm.pt — final fine-tuned checkpoint
• tokenizer.json — BPE tokenizer
• model_config.json — architecture configuration
• training_summary.json — training statistics
• requirements.txt — runtime dependencies