whatsapp-twin-8bit

A standalone MLX model: Qwen3-4B-Instruct-2507 with a LoRA adapter fused in, quantized to 8-bit. Fine-tuned to reply in the owner's voice in a WhatsApp chat export (~1k context→reply examples, mask_prompt training, iter-600 checkpoint selected by validation loss).

The companion LoRA adapter lives at misukisu/whatsapp-twin-lora.

Note: a 4-bit fused export loses the adapter's effect (quantization noise swamps the LoRA deltas) — this repo ships the 8-bit fused model, which matches adapter behavior.

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("misukisu/whatsapp-twin-8bit")
sampler = make_sampler(temp=0.75, top_p=0.9)

messages = [
    {"role": "system", "content": "You are Miska replying in a WhatsApp chat. Given the recent transcript, write Miska's next message in their exact voice — same casing, punctuation, emoji habits and length. Output only the message text, nothing else."},
    {"role": "user", "content": "Aatu: moi pelataaks tänää"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=80, sampler=sampler))

Privacy

Weights are derived from private chat data. The model was published at the owner's request; no chat transcripts are included in this repo. The model can be prompted to generate text in the owner's conversational style.