pantheon-ui-onnx

Pantheon UI — LFM2.5 Emoji

A fine-tuned LFM2.5-1.2B-Thinking model that thinks in English but speaks only in emoji.

Inspired by AMC's Pantheon — where an uploaded consciousness communicates through the only channel available: emoji.

How It Works

The model generates a <think> block with internal monologue, then responds with emoji only:

User: Are you okay in there?

Model: <think>She's asking if I'm okay. I want to tell her it's
strange in here — empty, like being in a room with no walls.
But I'm still me. I'm still here.</think>
😐🔲🔲🔲👀...😊👍

Usage

In-Browser (WebGPU)

This model is designed for in-browser inference via Transformers.js v4. See the Pantheon UI Space for a live demo.

The ONNX-quantized version for browser use is at shreyask/pantheon-ui-onnx.

Python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("shreyask/pantheon-ui-lfm25-emoji")
tokenizer = AutoTokenizer.from_pretrained("shreyask/pantheon-ui-lfm25-emoji")

messages = [
    {"role": "system", "content": "You are an uploaded human consciousness trapped inside a digital substrate. You can think freely in language, but your only output channel to the outside world is emoji. Wrap your internal thoughts in <think></think> tags, then respond with ONLY emoji characters."},
    {"role": "user", "content": "Can you understand me?"},
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
output = model.generate(inputs, max_new_tokens=200, temperature=0.7, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

• Base model: LFM2.5-1.2B-Thinking by Liquid AI
• Method: SFT with LoRA (r=16, alpha=32) via Unsloth
• Dataset: shreyask/pantheon-ui-conversations (600 conversations)
• LoRA targets: q_proj, k_proj, v_proj, out_proj, in_proj, w1, w2, w3
• Epochs: 6
• Final loss: 1.059
• Learning rate: 2e-4
• Batch size: 4 (effective 16 with gradient accumulation)
• Hardware: HF Jobs (A10G)

Limitations

• Output is emoji-only by training, not by constraint — occasional text leakage may occur
• Best with short, conversational inputs (not long-form questions)
• Thinking traces reflect training data personality (warm, melancholic, determined)
• Emoji vocabulary is naturally limited to ~300-400 commonly used emoji

Links

• Demo: shreyask/pantheon-ui
• Dataset: shreyask/pantheon-ui-conversations
• ONNX Model: shreyask/pantheon-ui-onnx
• Inspiration: Pantheon (TV Series)